Power load prediction method and device for data center

Abstract

The invention discloses a power load prediction method and device for a data center, and relates to the technical field of power load prediction. According to the method, the historical load data is firstly divided into the plurality of data sets, and then independent decomposition processing and feature extraction are performed on each data set, so that the mutual influence among the data sets is avoided, the data leakage risk among the data sets is effectively prevented, the independence and accuracy of the data sets are ensured, and the accuracy of the data sets is improved. The accuracy of each feature extracted based on the data set is ensured; by mining the internal association between the historical energy consumption attribute data of the data center and the load characteristics, the accuracy of the power load prediction result obtained based on the historical load data is improved; through independent decomposition of the data sets, independent feature extraction of the data sets and independent prediction of the data sets, information transverse flow between the data sets is avoided, and thus the power load prediction accuracy of the data center is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power load prediction, and in particular relates to a method and device for predicting power load in a data center. Background Art

[0002] With the booming development of information technologies such as big data, cloud computing, and artificial intelligence, the scale and number of data centers as infrastructure have continued to increase, bringing new challenges: the energy consumption of data centers. In order to provide stable and high-quality services, the servers in the data centers need to run around the clock, resulting in huge power consumption in the data centers.

[0003] Different from the load characteristics of ordinary commercial or residential buildings, the power demand of data centers presents the characteristics of continuous high load. Since they need to operate uninterruptedly around the clock, their power consumption is basically maintained at a high and stable level. Moreover, they present complex and changeable characteristics. The existing power load forecasting technology for data centers ignores the complexity of power load in data centers, resulting in inaccurate load forecasting through large models such as neural network models. Therefore, a new method is urgently needed to solve the defects of existing technologies. Summary of the invention

[0004] The present invention aims to provide a method and device for power load prediction of a data center to solve the above-mentioned technical problems, decompose and extract features of historical power load data of the data center, realize power load prediction based on historical power load data, and improve the accuracy of power load prediction for the data center.

[0005] In order to solve the above technical problems, an embodiment of the present invention provides a method for predicting power load of a data center, comprising:

[0006] Acquire historical load data, and divide the historical load data into several data sets;

[0007] Decomposing each of the data sets according to a preset decomposition technique to obtain a function set of each of the data sets;

[0008] According to the historical energy consumption attribute data, in combination with the function set, a feature vector corresponding to each function set is obtained;

[0009] According to a preset feature learning method, the feature vectors are combined to obtain a sequence component of each feature vector;

[0010] According to a preset reconstruction method, the sequence components are combined to obtain a power load forecast result.

[0011] It can be understood that, compared with the prior art, the present invention avoids mutual influence between data sets by first dividing the historical load data into multiple data sets, and then independently decomposing, processing and extracting features for each data set, thereby effectively preventing the risk of data leakage between data sets, thereby ensuring the independence and accuracy of the data sets, and further ensuring the accuracy of each feature extracted based on the data set; by mining the intrinsic relationship between the historical energy consumption attribute data and load characteristics of the data center, the accuracy of the power load forecasting results based on the historical load data is improved; by independently decomposing the data sets, independently extracting features from the data sets and independently predicting the data sets, the horizontal flow of information between the data sets is avoided, thereby improving the accuracy of the power load forecasting of the data center.

[0012] As a preferred solution, the method of decomposing each data set according to a preset decomposition technology to obtain a function set of each data set specifically includes:

[0013] Formulate constrained variational optimization problems;

[0014] According to a preset augmented Lagrangian function, the constrained variational optimization problem is changed to obtain an unconstrained variational optimization problem;

[0015] According to the unconstrained variational optimization problem, in combination with a preset decomposition algorithm and a preset number of decomposition layers, each of the data sets is decomposed to obtain a function set of each of the data sets.

[0016] This preferred solution decomposes the data set through a preset decomposition algorithm and a preset number of decomposition layers, combined with an unconstrained variational optimization problem, thereby achieving independent decomposition of the data set, so that the function set of each data set has the same number of decomposition layers, avoiding interference with the data set due to differences in the number of decomposition layers, thereby ensuring the accuracy of the decomposition of the data set, and further ensuring the accuracy of the power load forecasting results based on historical load data.

[0017] As a preferred solution, the characteristic vector corresponding to each function set is obtained based on the historical energy consumption attribute data and combined with the function set, specifically including:

[0018] Acquire historical energy consumption attribute data, wherein the historical energy consumption attribute data includes: outdoor environment temperature, humidity and power supply efficiency;

[0019] According to the outdoor environment temperature, the humidity, the power efficiency, and each of the function sets, a plurality of input matrices are obtained;

[0020] Convolution pooling is performed on each of the input matrices to obtain a feature vector corresponding to each function set.

[0021] This preferred solution realizes feature extraction by constructing an input matrix and performing convolution pooling; it deeply analyzes the intrinsic relationship between historical energy consumption attribute data and historical load, thereby improving the accuracy of power load forecasting results based on historical load data.

[0022] As a preferred solution, the sequence components of each feature vector are obtained by combining the feature vector according to the preset feature learning method, which specifically includes:

[0023] According to a preset forgetting function, information of each of the feature vectors is discarded to obtain a plurality of first feature vectors;

[0024] According to a preset update function, information screening is performed on each of the first feature vectors to obtain an information vector corresponding to each of the first feature vectors;

[0025] According to a preset calculation function, the information vector corresponding to each first eigenvector is updated to obtain the sequence component of each eigenvector.

[0026] This preferred solution obtains the sequence components by extracting the time series features in the feature vector, and then reconstructs the sequence components to obtain the prediction results. By extracting the time series features, the relationship between load and time can be fully explored, thereby deepening the understanding of historical load data and ensuring the accuracy of power load prediction results based on historical load data.

[0027] As a preferred solution, the power load forecasting result is obtained by combining the sequence components according to the preset reconstruction method, which specifically includes:

[0028] According to a preset reconstruction method, the predicted value of each sequence component is integrated to obtain a power load prediction result.

[0029] This preferred solution can obtain the power load forecast result by integrating the sequence components, so that the power load forecast result can be obtained based on the historical power load data.

[0030] Accordingly, an embodiment of the present invention provides a power load prediction device for a data center, comprising: a data decomposition module, a feature extraction module, a sequence component acquisition module and a power load prediction module;

[0031] The data partitioning module is used to obtain historical load data and partition the historical load data to obtain several data sets;

[0032] The data decomposition module is used to decompose each of the data sets according to a preset decomposition technology to obtain a function set of each of the data sets;

[0033] The feature extraction module is used to obtain a feature vector corresponding to each function set based on the historical energy consumption attribute data and the function set;

[0034] The sequence component acquisition module is used to obtain the sequence component of each feature vector in combination with the feature vector according to a preset feature learning method;

[0035] The power load prediction module is used to obtain a power load prediction result according to a preset reconstruction method in combination with the sequence components.

[0036] As a preferred solution, the data decomposition module includes: a data decomposition unit;

[0037] The data decomposition unit is used to obtain historical load data and a preset augmented Lagrangian function.

[0038] Dividing the historical load data to obtain a plurality of data sets;

[0039] Formulate constrained variational optimization problems;

[0040] According to a preset augmented Lagrangian function, the constrained variational optimization problem is changed to obtain an unconstrained variational optimization problem;

[0041] According to the unconstrained variational optimization problem, combined with a preset decomposition algorithm, each of the data sets is decomposed to obtain a function set of each of the data sets.

[0042] As a preferred solution, the feature extraction module includes: a feature extraction unit;

[0043] The feature extraction unit is used to obtain historical energy consumption attribute data, wherein the historical energy consumption attribute data includes: outdoor environment temperature, humidity and power supply efficiency;

[0044] According to the outdoor environment temperature, the humidity, the power efficiency, and each of the function sets, a plurality of input matrices are obtained;

[0045] Convolution pooling is performed on each of the input matrices to obtain a feature vector corresponding to each function set.

[0046] As a preferred solution, the sequence component acquisition module includes: a sequence component acquisition unit;

[0047] The sequence component acquisition unit is used to discard information of each of the feature vectors according to a preset forgetting function to obtain a plurality of first feature vectors;

[0048] According to a preset update function, information screening is performed on each of the first feature vectors to obtain an information vector corresponding to each of the first feature vectors;

[0049] According to a preset calculation function, the information vector corresponding to each first eigenvector is updated to obtain the sequence component of each eigenvector.

[0050] As a preferred solution, the power load prediction module includes: a power load prediction unit;

[0051] The power load prediction unit is used to integrate the prediction value of each sequence component according to a preset reconstruction method to obtain a power load prediction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 : A flowchart of a method for predicting power load in a data center provided by an embodiment of the present invention;

[0053] Figure 2 : A schematic diagram of the architecture of a VMD-CNN-LSTM model provided in an embodiment of the present invention;

[0054] Figure 3 : A structural diagram of an LSTM network provided in an embodiment of the present invention;

[0055] Figure 4 : A schematic diagram of the structure of a power load prediction device for a data center provided by an embodiment of the present invention;

[0056] Among them, 201: data division module; 202: data decomposition module; 203: feature extraction module; 204: sequence component acquisition module; 205: power load prediction module. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] With the booming development of information technologies such as big data, cloud computing, and artificial intelligence, the scale and number of data centers (DCs) as infrastructure have continued to increase, bringing new challenges - the energy consumption of data centers. In order to provide users with stable and high-quality services, the servers in data centers need to run around the clock, resulting in huge power consumption in data centers. Therefore, how to improve the energy efficiency of data centers is an urgent problem that needs to be solved for both the power industry and data center service providers.

[0059] The load characteristics of power data centers are reflected in their unique power demand patterns, which directly affect energy management strategies, cost control and the stability of system operation. Different from the load characteristics of ordinary commercial or residential buildings, the power demand of data centers presents the characteristics of continuous high load. Due to the need for uninterrupted operation around the clock, its power consumption is basically maintained at a high and stable level. However, this load is not constant, but nonlinear and non-stationary. It is affected by multiple factors such as server workload, data transmission volume and cooling system requirements, showing complex and changeable characteristics. In addition, the internal equipment of data centers is dense and the power density is extremely high, requiring the power supply system to withstand high loads and ensure power supply quality. In order to improve energy efficiency, data centers use cooling technologies such as cold and hot channel isolation, and are equipped with redundant power supply and backup power supply to ensure business continuity and data security, but this also increases the complexity of power load. The energy efficiency indicator PUE (power usage efficiency) has become the key to measuring the effectiveness of data center energy management, reflecting the ratio of total energy consumption of data centers to energy consumption of IT equipment. Optimizing PUE has become the goal pursued by the industry. Seasonal changes also affect the power demand of data centers, especially the power consumption of cooling systems will change with external temperature fluctuations.

[0060] Accurately predicting the energy consumption indicators of data centers can not only provide an important basis for effectively reducing energy consumption in data centers, but also improve the effective management of electric energy and optimize the power distribution system of data centers. With the development of neural networks and machine learning, new intelligent prediction methods for power loads have been further studied. In order to take into account the temporal and nonlinear relationship of data, recurrent neural networks (RNN) and convolutional neural networks (CNN) are applied to the load forecasting model of power systems. The model can predict longer historical data sequences and accurately map the complex nonlinear relationship between load influencing factors and loads. In order to solve the problem that gradients of RNN and CNN are prone to vanishing and exploding in long sequence training, a load forecasting model based on long short-term memory network (LSTM) was born. In addition, based on the empirical mode decomposition method (EMD), the load sequence is decomposed into multiple smooth sequences, combined with the LSTM method to predict one by one, and finally the sequence is reorganized. The results show that decomposing and reorganizing the load sequence can obtain a more accurate prediction sequence. However, although the feature of automatically determining the number of decomposed intrinsic mode functions (IMFs) based on the local characteristics of the signal is flexible, it may also lead to inadvertent data leakage during the model building process.

[0061] For example, in the existing neural network-based power load forecasting technology for data centers, the EMD-CNN-LSTM model is used for forecasting, but this method has many problems. The first problem is that there is a risk of data leakage in the EMD (empirical mode decomposition) stage. Due to the characteristic of EMD that automatically determines the number of IMFs, the IMFs between different data sets are inconsistent, which may leak future signal features in the validation set and the test set to the training set, affecting the model's accurate prediction ability; the second problem is that there is a risk of information leakage during the training of the EMD-CNN-LSTM model; the third problem is that the model's complex structure leads to low prediction ability and low versatility; the fourth problem is that there is a leakage risk in the hyperparameter tuning of the EMD-CNN-LSTM model.

[0062] The existing technology also uses CNN-LSTM technology for prediction, but if the CNN-LSTM model is directly applied to process time series data without using EMD, problems such as high-frequency noise interference and limited capture of long-term and short-term dependencies will occur, that is, the ability to process specific types of complex signals is lacking and inefficient. Among them, the first problem is that the CNN-LSTM model lacks a preprocessing step to filter or weaken the noise, which affects the CNN-LSTM model's learning of feature signals, that is, there will be a problem of high-frequency noise interference; the second problem is that only using the CNN-LSTM model has limitations in capturing long-term and short-term dependencies; the third problem is that only using the CNN-LSTM model is not accurate and targeted enough in feature expression; the fourth problem is that directly inputting the original data into the CNN-LSTM model will make the model too complex and there is a risk of overfitting.

[0063] As mentioned above, EMD, as an adaptive decomposition technology based on local signal features, is good at decomposing nonlinear and non-stationary signals into a series of intrinsic mode functions (IMFs). Each IMF captures the vibration state within a specific frequency range and reflects the different frequency components of the signal. Although the adaptive characteristics of EMD give it unique flexibility in processing complex signals, in model building scenarios, especially in the construction of deep learning models, this feature leads to considerations of data leakage. Specifically, the data set needs to be divided into training, validation, and test subsets to ensure the generalization ability of the model. When EMD is applied to these subsets alone, the mechanism of automatically determining the number of IMFs may lead to inconsistent IMFs between subsets, thereby affecting the fairness and effectiveness of model training.

[0064] Embodiment 1

[0065] Please refer to Figure 1, is a step flow chart of a method for predicting power load of a data center provided in an embodiment of the present invention, including steps S101 to S105.

[0066] Please refer to Figure 2 , which is a schematic diagram of the architecture of a VMD-CNN-LSTM model provided in an embodiment of the present invention; it should be noted that steps S101 to S105 described in the embodiment of the present invention can not only be based on Figure 2 The VMD-CNN-LSTM model shown can also be implemented by other neural network models. The following embodiments of the present invention further illustrate steps S101 to S105 in conjunction with the VMD-CNN-LSTM model.

[0067] like Figure 2 As shown, the VMD-CNN-LSTM model includes: an input layer, a CNN layer, an LSTM layer and an output layer; the input layer is used to process data, the CNN layer is used to perform convolution pooling on the data input to the CNN layer, the LSTM layer is used to capture long-term and short-term dependencies, and the output layer is used to reconstruct the component prediction values ​​to obtain prediction values, as well as to perform error evaluation analysis.

[0068] Step S101: Acquire historical load data, and divide the historical load data into several data sets.

[0069] In certain implementations of this embodiment, Figure 2 As shown, historical load data of a data center is obtained, and the historical load data is divided into a training set, a validation set, and a test set in an input layer.

[0070] Step S102: Decompose each of the data sets according to a preset decomposition technique to obtain a function set of each of the data sets.

[0071] In this embodiment, the decomposition of each data set according to a preset decomposition technology to obtain a function set of each data set specifically includes:

[0072] Formulate constrained variational optimization problems;

[0073] According to a preset augmented Lagrangian function, the constrained variational optimization problem is changed to obtain an unconstrained variational optimization problem;

[0074] According to the unconstrained variational optimization problem, in combination with a preset decomposition algorithm and a preset number of decomposition layers, each of the data sets is decomposed to obtain a function set of each of the data sets.

[0075] In certain implementations of this embodiment, Figure 2As shown, the preset decomposition technology is VMD technology (Variational Mode Decomposition).

[0076] In an optional embodiment, the constrained variational optimization problem includes:

[0077]

[0078] Among them, u k (t) indicates that any original input signal f(t) can be decomposed into k modal components u k ;u k (t)={u1,u2,…,u k} represents the set of modal components; w k represents the center frequency of the kth mode after the original input signal f(t) is decomposed; {w k}={w1,w2,...,w k} represents a set of center frequencies corresponding to the signal decomposition, and δ(t) represents the unit pulse function.

[0079] In an optional embodiment, the preset augmented Lagrangian function (augmented Lagrange) includes:

[0080]

[0081] Among them, λ(t) represents the Lagrange multiplication operator and α represents the quadratic penalty term.

[0082] In an optional embodiment, the unconstrained variational optimization problem is combined with a preset decomposition algorithm and a preset number of decomposition layers to decompose each of the data sets to obtain a function set of each of the data sets, specifically including:

[0083] Use the alternate direction method of multiplication algorithm (ADMM) to solve the saddle point of formula (2) and get the result u k and w k , a set of modal components can be represented by the estimated values ​​of the solution in the frequency domain, and the function set represented by formula (3) and formula (4) is obtained:

[0084]

[0085] Where n represents the number of iterations; and They represent f(t), u i The Fourier transform expressions of (t) and λ(t) are: Is the remaining part The Wiener filter result is, is the center frequency.

[0086] Based on the above implementation, VMD technology is used to adaptively decompose the original time series data into multiple intrinsic mode functions (IMFs) according to the intrinsic frequency characteristics of the signal. This process can not only separate the signal components of different frequency bands, but also effectively reduce the impact of sequence fluctuations on subsequent prediction models and improve the interpretability of the signal.

[0087] It should be noted that the main work of VMD technology is to decompose a signal containing multiple mixed sub-components into multiple single finite bandwidth sub-modal components. In addition, VMD technology can also obtain the original input signal through reverse work.

[0088] It should be noted that by presetting the number of modalities through VMD, a fixed number of IMF decomposition layers is designed for the entire data set before processing the training set, validation set, and test set, ensuring that all data subsets undergo the same operation during the data preprocessing stage, avoiding potential data leakage problems caused by automatically determining the number of IMFs. By controlling the number of decomposition layers, the consistency between data sets is maintained, the independence of the validation set and test set information is protected, and the fairness of model evaluation is ensured.

[0089] Step S103: According to the historical energy consumption attribute data, in combination with the function set, a feature vector corresponding to each function set is obtained.

[0090] In this embodiment, the characteristic vector corresponding to each function set is obtained based on the historical energy consumption attribute data and combined with the function set, specifically including:

[0091] Acquire historical energy consumption attribute data, wherein the historical energy consumption attribute data includes: outdoor environment temperature, humidity and power supply efficiency;

[0092] According to the outdoor environment temperature, the humidity, the power efficiency, and each of the function sets, a plurality of input matrices are obtained;

[0093] Convolution pooling is performed on each of the input matrices to obtain a feature vector corresponding to each function set.

[0094] In an alternative embodiment, if Figure 2 As shown, in the input layer, the historical energy consumption attribute data of the data center can also be extracted to obtain the outdoor environment temperature, humidity and power usage efficiency (PUE) of the data center.

[0095] In certain implementations of this embodiment, Figure 2As shown, the CNN layer includes: a first layer of one-dimensional convolution, a second layer of one-dimensional convolution, two maximum poolings and one Dropout;

[0096] In an alternative embodiment, if Figure 2 As shown, in the CNN layer, the outdoor ambient temperature, humidity, power efficiency and function set are combined to obtain several input matrices; the input matrices are convolved and pooled through the first layer of one-dimensional convolution, the second layer of one-dimensional convolution, two maximum pooling and Dropout to obtain the feature vector corresponding to each function set.

[0097] Based on the above implementation methods, by constructing an input matrix, while maintaining the signal frequency characteristics, the impact of environmental factors on the load is incorporated to provide richer input information for the CNN layer; by performing a pooling operation after convolution, the spatial dimension of the data can be reduced, the amount of calculation can be reduced, and the characteristics of the data can be maintained; through Dropout, in the convolution pooling process of the CNN layer, the outputs of some neurons are randomly discarded, thereby forcing the CNN layer to have a stronger feature representation, thereby improving the processing ability of the CNN layer on unseen data; through convolution pooling operations, the intrinsic correlation between the outdoor ambient temperature, humidity and power utilization efficiency of the data center and the load sequence is deeply analyzed; the data of the training set, validation set and test set are independently processed according to the principle of independence and closure, ensuring that there is no lateral flow of information during the feature extraction process, and effectively isolating the potential risk of information leakage between different data sets.

[0098] Step S104: according to a preset feature learning method, the feature vectors are combined to obtain the sequence components of each feature vector.

[0099] In this embodiment, the sequence components of each feature vector are obtained by combining the feature vector according to the preset feature learning method, which specifically includes:

[0100] According to a preset forgetting function, information of each of the feature vectors is discarded to obtain a plurality of first feature vectors;

[0101] According to a preset update function, information screening is performed on each of the first feature vectors to obtain an information vector corresponding to each of the first feature vectors;

[0102] According to a preset calculation function, the information vector corresponding to each first eigenvector is updated to obtain the sequence component of each eigenvector.

[0103] In an alternative embodiment, please refer to Figure 3, is a structural diagram of an LSTM network provided by an embodiment of the present invention, wherein the LSTM network (i.e., LSTM layer) includes: a forget gate, an input gate, and an output gate; the feature vector of the CNN layer is input to the LSTM layer, and passes through the forget gate, input gate, and output gate in sequence to obtain the time series features in the feature vector.

[0104] In an optional embodiment, the preset forgetting function is a forgetting gate, the preset updating function is an input gate, and the preset calculation function is an output gate.

[0105] It should be noted that if Figure 2 and Figure 3 As shown, the forget gate is used to filter out the memory state, and the activation state of the information in the forget gate can be calculated by the sigmoid function; the input gate is used to determine the information to be stored, and the output gate is used to update the information.

[0106] In an alternative embodiment, if Figure 2 and Figure 3 As shown, the function of the forget gate is:

[0107] f t =σ(W f ·[h t-1 ,x t ]+b f ) (5),

[0108] The function of the input gate is:

[0109] i t =σ(W i ·[h t-1 ,x t ]+b i ) (6),

[0110]

[0111] The memory unit of the LSTM network can be expressed as:

[0112]

[0113] The function of the output gate is:

[0114] o t =σ(W o [h t-1 ,x t ]+b o ) (9),

[0115] h t =o t *tanh (c t) (10),

[0116] Among them, W f , W i , W c and W o is the weight matrix, vector b f 、b i 、b c and b o is the bias vector, x t represents the input vector, h t Represents the hidden layer state quantity, σ∈[0,1], and the size of σ indicates the degree to which the information retains the previous memory state.

[0117] It should be noted that if Figure 2 As shown in the above function, the first step is as shown in formula (5), using f t Represents the forget gate, which is used to decide which information to choose to get rid of the memory state. The activation of the forget gate can be calculated by the sigmoid function;

[0118] The second step is to determine what new information is stored in the cell state, as shown in formulas (6) and (7). First, the "S-type" activation function in the input gate selects the past information, and then the tanh function is used to process the information to generate the vector This vector is updated in the next state;

[0119] The third step is to update the memory unit as shown in formula (8);

[0120] The fourth step is as shown in formulas (9) and (10). The information is output after retaining the selected memory through the output gate. First, the sigmoid function is used as the output gate to filter the memory state. Then, the memory state is placed through the tanh function and multiplied by the output o. t To calculate the amount of information required.

[0121] Based on the above implementation method, the LSTM network uses its unique gating mechanism to memorize long-term dependencies and efficiently extract time series features of sequence summaries; by only accepting the feature vectors output by the CNN layer and following the data isolation principle, the purity and reliability of LSTM during the working process are ensured.

[0122] Step S105: According to a preset reconstruction method, the sequence components are combined to obtain a power load prediction result.

[0123] In this embodiment, the power load forecasting result is obtained by combining the sequence components according to the preset reconstruction method, which specifically includes:

[0124] According to a preset reconstruction method, the predicted value of each sequence component is integrated to obtain a power load prediction result.

[0125] In an alternative embodiment, if Figure 2 As shown, in the output layer, the predicted values ​​of all sequence components output by the LSTM layer are integrated, and the final power load forecasting result is reconstructed according to a preset combination or weighted summation method.

[0126] Based on the above implementation, by reconstructing at the output layer to independently verify and correct the training set, validation set and test set, following the isolation and verification mechanism, the independence and accuracy of the prediction of each data set is ensured, and data leakage is avoided.

[0127] In an optional embodiment, after obtaining the power load forecast result, the forecast result is compared with the actual load data, and error analysis and evaluation are performed, so as to adjust and optimize the VMD-CNN-LSTM model according to the evaluation result.

[0128] It should be noted that the error analysis method described in the embodiment of the present invention includes: relative error (RelativeError, RE), mean square error (Mean Square Error, MSE), mean absolute error (Mean Absolute Error, MAE), root mean square error (Root Mean Square Error, RMSE) and mean absolute percentage error (Mean Absolute Percentage Error, MAPE).

[0129] Among them, the function of the error analysis method is as follows:

[0130]

[0131] The function of the mean square error is as follows:

[0132]

[0133] The function of the mean absolute error is as follows:

[0134]

[0135] The function of the root mean square error is as follows:

[0136]

[0137] The function of the mean absolute percentage error is as follows:

[0138]

[0139] In the above formulas (11) to (15), y i is the true value, is the predicted value.

[0140] This embodiment first divides the historical load data into multiple data sets, and then independently decomposes and processes each data set and extracts features, thereby avoiding mutual influence between data sets, thereby effectively preventing the risk of data leakage between data sets, thereby ensuring the independence and accuracy of the data sets, and further ensuring the accuracy of each feature extracted based on the data set; by mining the intrinsic relationship between the historical energy consumption attribute data and load characteristics of the data center, the accuracy of the power load prediction results based on the historical load data is improved; by independently decomposing the data sets, independently extracting features from the data sets, and independently predicting the data sets, the horizontal flow of information between the data sets is avoided, thereby improving the accuracy of the power load prediction of the data center.

[0141] Embodiment 2

[0142] Please refer to Figure 4 , which is a structural schematic diagram of a power load prediction device for a data center provided in an embodiment of the present invention, comprising: a data partitioning module 201, a data decomposition module 202, a feature extraction module 203, a sequence component acquisition module 204 and a power load prediction module 205.

[0143] The data partitioning module 201 is used to obtain historical load data and partition the historical load data to obtain several data sets.

[0144] The data decomposition module 202 is used to decompose each of the data sets according to a preset decomposition technology to obtain a function set of each of the data sets.

[0145] In this embodiment, the data decomposition module 202 includes: a data decomposition unit;

[0146] The data decomposition unit is used to construct a constrained variational optimization problem;

[0147] According to a preset augmented Lagrangian function, the constrained variational optimization problem is changed to obtain an unconstrained variational optimization problem;

[0148] According to the unconstrained variational optimization problem, in combination with a preset decomposition algorithm and a preset number of decomposition layers, each of the data sets is decomposed to obtain a function set of each of the data sets.

[0149] The feature extraction module 203 is used to obtain a feature vector corresponding to each function set based on the historical energy consumption attribute data and in combination with the function set.

[0150] In this embodiment, the feature extraction module 203 includes: a feature extraction unit;

[0151] The feature extraction unit is used to obtain historical energy consumption attribute data, wherein the historical energy consumption attribute data includes: outdoor environment temperature, humidity and power supply efficiency;

[0152] According to the outdoor environment temperature, the humidity, the power efficiency, and each of the function sets, a plurality of input matrices are obtained;

[0153] Convolution pooling is performed on each of the input matrices to obtain a feature vector corresponding to each function set.

[0154] The sequence component acquisition module 204 is used to obtain the sequence component of each feature vector in combination with the feature vector according to a preset feature learning method.

[0155] In this embodiment, the sequence component acquisition module 204 includes: a sequence component acquisition unit;

[0156] The sequence component acquisition unit is used to discard information of each of the feature vectors according to a preset forgetting function to obtain a plurality of first feature vectors;

[0157] According to a preset update function, information screening is performed on each of the first feature vectors to obtain an information vector corresponding to each of the first feature vectors;

[0158] According to a preset calculation function, the information vector corresponding to each first eigenvector is updated to obtain the sequence component of each eigenvector.

[0159] The power load prediction module 205 is used to obtain a power load prediction result according to a preset reconstruction method and in combination with the sequence components.

[0160] In this embodiment, the power load prediction module 205 includes: a power load prediction unit;

[0161] The power load prediction unit is used to integrate the prediction value of each sequence component according to a preset reconstruction method to obtain a power load prediction result.

[0162] This embodiment first divides the historical load data into multiple data sets, and then independently decomposes and processes each data set and extracts features, thereby avoiding mutual influence between data sets, thereby effectively preventing the risk of data leakage between data sets, thereby ensuring the independence and accuracy of the data sets, and further ensuring the accuracy of each feature extracted based on the data set; by mining the intrinsic relationship between the historical energy consumption attribute data and load characteristics of the data center, the accuracy of the power load prediction results based on the historical load data is improved; by independently decomposing the data sets, independently extracting features from the data sets, and independently predicting the data sets, the horizontal flow of information between the data sets is avoided, thereby improving the accuracy of the power load prediction of the data center.

[0163] In summary, the embodiments of the present invention avoid mutual influence between data sets by first dividing the historical load data into multiple data sets, and then independently decomposing, processing and extracting features for each data set, thereby effectively preventing the risk of data leakage between data sets, thereby ensuring the independence and accuracy of the data sets, and further ensuring the accuracy of each feature extracted based on the data set; by mining the intrinsic relationship between the historical energy consumption attribute data and load characteristics of the data center, the accuracy of the power load forecasting results obtained based on the historical load data is improved; by independently decomposing the data sets, independently extracting features of the data sets and independently predicting the data sets, the horizontal flow of information between the data sets is avoided, thereby improving the accuracy of the power load forecasting of the data center.

[0164] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting power load in a data center. It is characterized in that include: Acquire historical load data, and divide the historical load data into several data sets; Decomposing each of the data sets according to a preset decomposition technique to obtain a function set of each of the data sets; According to the historical energy consumption attribute data, combined with the function set, a feature vector corresponding to each function set is obtained; According to a preset feature learning method, the feature vectors are combined to obtain a sequence component of each feature vector; According to a preset reconstruction method, the sequence components are combined to obtain a power load forecast result.

2. A method for predicting power load in a data center according to claim 1, It is characterized in that Decomposing each of the data sets according to a preset decomposition technique to obtain a function set of each of the data sets specifically includes: Formulate constrained variational optimization problems; According to a preset augmented Lagrangian function, the constrained variational optimization problem is changed to obtain an unconstrained variational optimization problem; According to the unconstrained variational optimization problem, in combination with a preset decomposition algorithm and a preset number of decomposition layers, each of the data sets is decomposed to obtain a function set of each of the data sets.

3. A method for predicting power load in a data center according to claim 1, It is characterized in that The step of obtaining the characteristic vector corresponding to each function set based on the historical energy consumption attribute data and combining the function set specifically includes: Acquire historical energy consumption attribute data, wherein the historical energy consumption attribute data includes: outdoor environment temperature, humidity and power use efficiency; According to the outdoor environment temperature, the humidity, the power efficiency, and each of the function sets, a plurality of input matrices are obtained; Convolution pooling is performed on each of the input matrices to obtain a feature vector corresponding to each function set.

4. A method for predicting power load in a data center according to claim 1, It is characterized in that The step of obtaining the sequence component of each feature vector by combining the feature vector according to the preset feature learning method specifically includes: According to a preset forgetting function, information of each of the feature vectors is discarded to obtain a plurality of first feature vectors; According to a preset update function, information screening is performed on each of the first feature vectors to obtain an information vector corresponding to each of the first feature vectors; According to a preset calculation function, the information vector corresponding to each first eigenvector is updated to obtain the sequence component of each eigenvector.

5. A method for predicting power load in a data center according to any one of claims 1 to 4, It is characterized in that The method of reconstructing the power load according to the preset method and combining the sequence components to obtain the power load forecast result specifically includes: According to a preset reconstruction method, the predicted value of each sequence component is integrated to obtain a power load prediction result.

6. A power load prediction device for a data center, It is characterized in that include: Data partitioning module, data decomposition module, feature extraction module, sequence component acquisition module and power load forecasting module; The data partitioning module is used to obtain historical load data and partition the historical load data to obtain several data sets; The data decomposition module is used to decompose each of the data sets according to a preset decomposition technology to obtain a function set of each of the data sets; The feature extraction module is used to obtain a feature vector corresponding to each function set based on the historical energy consumption attribute data and the function set; The sequence component acquisition module is used to obtain the sequence component of each feature vector in combination with the feature vector according to a preset feature learning method; The power load prediction module is used to obtain a power load prediction result according to a preset reconstruction method in combination with the sequence components.

7. A data center power load prediction device as claimed in claim 6, It is characterized in that The data decomposition module comprises: a data decomposition unit; The data decomposition unit is used to construct a constrained variational optimization problem; According to a preset augmented Lagrangian function, the constrained variational optimization problem is changed to obtain an unconstrained variational optimization problem; According to the unconstrained variational optimization problem, in combination with a preset decomposition algorithm and a preset number of decomposition layers, each of the data sets is decomposed to obtain a function set of each of the data sets.

8. The power load prediction device for a data center according to claim 6, It is characterized in that The feature extraction module comprises: a feature extraction unit; The feature extraction unit is used to obtain historical energy consumption attribute data, wherein the historical energy consumption attribute data includes: outdoor environment temperature, humidity and power supply efficiency; According to the outdoor environment temperature, the humidity, the power efficiency, and each of the function sets, a plurality of input matrices are obtained; Convolution pooling is performed on each of the input matrices to obtain a feature vector corresponding to each function set.

9. The power load prediction device for a data center according to claim 6, It is characterized in that The sequence component acquisition module comprises: a sequence component acquisition unit; The sequence component acquisition unit is used to discard information of each feature vector according to a preset forgetting function to obtain a plurality of first feature vectors; According to a preset update function, information screening is performed on each of the first feature vectors to obtain an information vector corresponding to each of the first feature vectors; According to a preset calculation function, the information vector corresponding to each first eigenvector is updated to obtain the sequence component of each eigenvector.

10. The power load prediction device for a data center according to any one of claims 6 to 9, It is characterized in that The power load prediction module comprises: a power load prediction unit; The power load prediction unit is used to integrate the prediction value of each sequence component according to a preset reconstruction method to obtain a power load prediction result.

Images