Electrical system load prediction method and system based on data center

Abstract

The invention discloses an electrical system load prediction method based on a data center, and the method comprises the steps: obtaining electrical parameters, environmental parameters and historical operation data of each subsystem through a data collection module, carrying out the normalization and time sequence reconstruction of the collected data, and forming standardized time sequence data; a combined neural network model comprising a convolutional neural network (CNN), an improved long-short term memory (LSTM) network and an attention mechanism is utilized to extract local fluctuation, mutation, frequency and statistical characteristics, and capture of data long-term dependence and key time sequence information is realized, so that an electrical load prediction value is output, and finally real-time regulation and control of a data center electrical system are realized. The method can effectively improve the prediction precision and response speed, and provides reliable technical support for the full life cycle management of the data center.

Description

Technical Field

[0001] This invention relates to the field of electrical system load forecasting technology, and specifically to a method and system for forecasting electrical system loads in data centers. Background Technology

[0002] With the rapid development of information technology and the widespread application of new technologies such as cloud computing, big data, and artificial intelligence, data centers, as important carriers of various computing and storage resources, are experiencing rapid growth in both scale and number. In the construction and operation of data centers, the electrical system, as a core supporting facility, plays a crucial role in the stability of the entire data center through its safe, reliable, and efficient operation. In recent years, with the increasing complexity of data center loads and the continuous changes in the operating environment, traditional electrical system load forecasting methods have gradually revealed many limitations. There is an urgent need for a new forecasting technology to achieve high-precision, real-time, and adaptive load forecasting, thereby providing strong support for the operation control and full lifecycle management of data centers.

[0003] Traditional load forecasting methods primarily rely on statistical approaches, such as time series analysis, regression analysis, and empirical model-based forecasting. These methods can achieve certain predictive results in scenarios with limited data volume and relatively simple patterns of change. However, with the increasing variety of data center equipment, the growing system complexity, and the large-scale, multi-dimensional, and non-linear characteristics of operational data, simply relying on traditional statistical models for forecasting can no longer meet the high-precision requirements of current data centers. First, traditional methods are highly sensitive to data noise and outliers, failing to fully uncover the deeper patterns hidden behind large datasets. Second, when faced with the complex coupling relationships between various subsystems within a data center, traditional forecasting methods struggle to capture the correlation between short-term fluctuations and long-term trends, thus affecting the accuracy of the forecast results. Finally, traditional methods often employ offline model training, lacking real-time online update capabilities, making it difficult to adapt to the dynamic changes in the data center operating environment.

[0004] In recent years, with the development of deep learning technology and artificial intelligence algorithms, deep neural networks, especially convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), have shown significant advantages in processing time-series data and extracting data features. CNNs can effectively extract local features from raw data, such as data fluctuations, abrupt changes, and frequency components, while LSTMs can capture long-term dependencies through memory mechanisms, solving the gradient vanishing problem in traditional recurrent neural networks. Furthermore, introducing attention mechanisms can dynamically allocate importance weights at each time step in the model, thereby further improving the sensitivity and accuracy of the prediction model to key features. The combined application of these technologies to form a synergistic neural network model holds promise for achieving fine-grained modeling and accurate prediction of data center electrical system loads.

[0005] Meanwhile, in actual operation, data centers generate a wealth of environmental parameters (such as temperature, humidity, air velocity, and air pressure) and historical operational data (such as equipment failure records, load curve data, and abnormal alarm data) in addition to electrical parameters. This data provides a rich source of information for load forecasting. However, these multi-source data suffer from inconsistent collection frequencies and varying data quality, requiring preprocessing operations such as data normalization and time-series data reconstruction to form standardized time-series data, providing high-quality input for subsequent deep neural network model training. Furthermore, with the dynamic changes in data center load, the predictions obtained from training traditional single neural network models are inaccurate and fail to leverage the strengths of each network model. Current technologies do not yet integrate and cascade multiple models to fully utilize their individual advantages in electrical system load forecasting.

[0006] Therefore, there is an urgent need for a data center electrical system load forecasting method and system that can combine neural network models to fully utilize electrical system data and adapt to real-time changes in load and environmental characteristics through combined models, thereby improving the accuracy and reliability of electrical system load forecasting. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, this invention proposes a method for load forecasting of data center electrical systems. This method acquires electrical parameters, environmental parameters, and historical operating data of each subsystem through a data acquisition module, and normalizes and reconstructs the acquired data to form standardized time-series data. Utilizing a combined neural network model incorporating convolutional neural networks (CNN), improved long short-term memory networks (LSTM), and attention mechanisms, it extracts local fluctuations, abrupt changes, frequencies, and statistical features to capture long-term data dependencies and key time-series information, thereby outputting predicted electrical load values ​​and ultimately enabling real-time control of the data center electrical system. This invention effectively improves prediction accuracy and response speed, providing reliable technical support for the entire lifecycle management of data centers.

[0008] A method for load forecasting based on data center electrical systems includes the following steps:

[0009] S1: The data acquisition module collects electrical parameters, environmental parameters, and historical operating data of each subsystem within the data center;

[0010] S2: Preprocess the collected data, including data normalization and time series data reconstruction to form standardized time series data;

[0011] S3: Feature extraction and load forecasting are performed on the standardized time-series data using a combined neural network model, wherein the combined neural network model includes:

[0012] CNN module: includes at least one convolutional layer and one pooling layer, wherein the convolutional layer uses convolutional kernels to extract local features from the normalized temporal data;

[0013] LSTM module: includes at least one improved long short-term memory network (LSTM) layer, which receives local features output by the CNN module and captures long-term temporal dependencies of the data, wherein the improved LSTM module includes gating units and residual connections to enhance memory performance;

[0014] Attention mechanism module: used to weight the temporal features output by the LSTM module;

[0015] S4: Output the predicted electrical load value based on the weighted feature vector;

[0016] S5: Implement real-time control of the data center electrical system based on the predicted results.

[0017] Preferably, the electrical parameters include: voltage values, current values, power factor, system frequency, harmonic content, active power and reactive power of the high-voltage system and the low-voltage system; the environmental parameters include: temperature, humidity, air velocity and air pressure in the data center computer room; the historical operating data includes: operating logs of each subsystem of the data center, load curve data, equipment fault records and abnormal alarm data.

[0018] Preferably, the data normalization includes applying min-max normalization or z-score normalization to the cleaned electrical and environmental parameters to map the data values ​​to a predetermined numerical range; the time-series data reconstruction includes sorting the data according to the collection timestamps of each data item, reconstructing the data with inconsistent sampling intervals into time-series data with a fixed sampling period using spline interpolation algorithm, and further reducing data noise by using moving average filtering to form standardized time-series data.

[0019] Preferably, the convolutional layers in the CNN module employ multiple 3×3 convolutional kernels of a fixed size, extracting local features from the temporal data using a sliding window approach, and the pooling layers employ max pooling to reduce data dimensionality and computational complexity. The local features include local fluctuation features, abrupt change features, local frequency components, and local statistical features. The local fluctuation features include the difference between the maximum and minimum values ​​of the data within the fixed sliding window, and the standard deviation of the data within the fixed sliding window. The abrupt change features include the maximum absolute difference between adjacent sampling points within the fixed sliding window. The local frequency components include the frequency component with the largest amplitude determined by applying a discrete Fourier transform to the data within the fixed sliding window. The local statistical features include the arithmetic mean and standard deviation within the local window.

[0020] Preferably, the improved LSTM module includes a residual connection structure, which is specifically implemented as follows:

[0021] First, in each LSTM unit, the input vector is processed through standard LSTM operations to obtain the hidden state output. Standard LSTM operations include gating computation, candidate memory state update, and hidden state computation.

[0022] Secondly, the residual connection structure will convert the input vector x of the current LSTM unit into a single input vector. t With output hidden state h t Perform element-wise addition to form the residual output y. t ;

[0023] y t =h t +f(x t )

[0024] Wherein, the function f(x)t ) is a linear mapping operation used to transform the input vector x t Mapped to h t A feature space with the same dimension; when the input dimension and the hidden state dimension are the same, f(x) t Take the identity mapping directly;

[0025] Finally, the residual connection structure not only connects the input and output within the same time step, but also allows the output of the previous LSTM module to be used as the input of the next LSTM module in a multi-layer LSTM structure. By transferring feature information across layers through residual connections, the ability to capture long-term dependent features is enhanced and the gradient vanishing problem is alleviated.

[0026] Preferably, the attention mechanism module includes:

[0027] A fully connected layer is used to generate an initial score from the LSTM output feature vector at each time step;

[0028] A softmax normalization layer converts the initial score into normalized attention weights;

[0029] The weighting operation multiplies the attention weights element-wise with the feature vector output by the LSTM to obtain a weighted feature vector, which is used as the final feature for load prediction.

[0030] The fully connected layer of the attention mechanism module has N neurons. The softmax normalization layer normalizes the N-dimensional preliminary score output by the fully connected layer. The normalization result is used as a weight vector and multiplied element-wise with the N-dimensional feature vector output by the LSTM, thereby achieving weighting of features at each time step.

[0031] The present invention also provides a load forecasting system based on a data center electrical system, comprising:

[0032] The data acquisition module collects electrical parameters, environmental parameters, and historical operating data of each subsystem within the data center.

[0033] The preprocessing module preprocesses the collected data, including data normalization and time-series data reconstruction to form standardized time-series data.

[0034] A combined neural network model computation module performs feature extraction and load prediction on the standardized time-series data. The combined neural network model includes:

[0035] CNN module: includes at least one convolutional layer and one pooling layer, wherein the convolutional layer uses convolutional kernels to extract local features from the normalized temporal data;

[0036] LSTM module: includes at least one improved long short-term memory network (LSTM) layer, which receives local features output by the CNN module and captures long-term temporal dependencies of the data, wherein the improved LSTM module includes gating units and residual connections to enhance memory performance;

[0037] Attention mechanism module: used to weight the temporal features output by the LSTM module;

[0038] The prediction output module outputs the electrical load prediction value based on the weighted feature vector;

[0039] The control module implements real-time control over the data center electrical system based on the predicted results.

[0040] Preferably, the electrical parameters include: voltage values, current values, power factor, system frequency, harmonic content, active power and reactive power of the high-voltage system and the low-voltage system; the environmental parameters include: temperature, humidity, air velocity and air pressure in the data center computer room; the historical operating data includes: operating logs of each subsystem of the data center, load curve data, equipment fault records and abnormal alarm data.

[0041] Preferably, the data normalization includes applying min-max normalization or z-score normalization to the cleaned electrical and environmental parameters to map the data values ​​to a predetermined numerical range; the time-series data reconstruction includes sorting the data according to the collection timestamps of each data item, reconstructing the data with inconsistent sampling intervals into time-series data with a fixed sampling period using spline interpolation algorithm, and further reducing data noise by using moving average filtering to form standardized time-series data.

[0042] Preferably, the convolutional layers in the CNN module employ multiple convolutional kernels of a fixed size of 3×3, extracting local features from the temporal data through a sliding window approach, and the pooling layers employ max pooling to reduce data dimensionality and computational complexity.

[0043] This invention provides a method and system for load forecasting of data center electrical systems, which can achieve the following beneficial technical effects:

[0044] 1. This invention achieves high-precision, real-time prediction of data center electrical system load by integrating technologies such as multi-source data acquisition, data normalization and time-series reconstruction, and combined neural network models. The prediction accuracy is high. It uses CNN to extract local fluctuations, mutations, frequencies and statistical features, combines an improved LSTM to capture long-term time-series dependencies, and further uses an attention mechanism for dynamic weighting. This invention can fully explore the deep-seated patterns in the data, effectively reduce noise interference, and improve the accuracy of load prediction.

[0045] 2. This invention makes full use of multidimensional data, comprehensively reflecting the operating status of the data center by collecting electrical parameters, environmental parameters and historical operating data, providing rich input for the prediction model, ensuring the comprehensiveness and accuracy of data processing and feature extraction, providing decision-making basis for accurate load forecasting and real-time control of the electrical system, helping to optimize power distribution, reduce equipment failure risk, further improve the overall operating efficiency and safety of the data center, and reduce operating costs.

[0046] 3. The improved LSTM module in this invention employs a residual connection structure. This structure effectively alleviates the gradient vanishing problem in deep networks by adding the input vector to the output of the LSTM unit element-wise, ensuring that gradients are fully propagated across multiple time steps and layers, thereby significantly enhancing the model's ability to capture long-term dependent features. This improvement not only enhances the training stability and convergence speed of the load prediction model in deep structures but also makes the model more accurate and robust when facing long-term complex time-series data from data center electrical systems. Furthermore, through residual connections, the model can better integrate low-level features and high-level semantic information, providing richer and more stable feature inputs for subsequent attention mechanism modules, thereby further improving the overall system's real-time prediction performance and adaptive control capabilities. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram illustrating the steps of a load forecasting method for data center electrical systems according to the present invention;

[0049] Figure 2 This invention relates to a load prediction system diagram based on a data center electrical system. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1:

[0052] In view of the aforementioned problems mentioned in the prior art, and in order to solve the above technical problems, as shown in the appendix. Figure 1 The method for load forecasting based on data center electrical systems includes the following steps:

[0053] S1: The data acquisition module collects electrical parameters, environmental parameters, and historical operating data of each subsystem within the data center. In one embodiment, electrical parameter acquisition is performed in both the high-voltage and low-voltage systems. High-precision voltage, current, and power factor sensors are installed in the high-voltage distribution room, and voltage (e.g., 380kV or 220kV), current, power factor, and system frequency are collected every second. Harmonic detection is performed in the low-voltage distribution room, where a harmonic analyzer monitors the harmonic content of the current in real time to ensure that the data records include harmonic component information for each frequency band. Active and reactive power measurement involves installing power meters on the main distribution cabinet and distribution cabinet to record the active and reactive power data of each circuit. This data helps in calculating the load balance. Environmental parameter collection includes the collection of computer room temperature and humidity. Temperature and humidity sensors are installed above and below the server racks in the data center to monitor the temperature (unit: °C) and humidity (unit: %RH) in real time. The data collection frequency is set to once per minute. Air velocity and air pressure are measured in the air-conditioned computer room. Air velocity sensors and air pressure sensors are installed to monitor the air circulation and ambient air pressure in the computer room, so as to determine the air conditioning load and ventilation effect. Historical operational data acquisition includes collecting operational log records. Each electrical device (such as distribution cabinets, UPS, and generators) is equipped with an intelligent monitoring system that records equipment status, fault alarms, and start / stop information in real time. The log data is stored in a structured manner in a central database. Load curve data acquisition involves the upper-level monitoring system periodically collecting load data from each circuit and generating historical load curves. The data retention period can be as long as several months or even years, providing historical data for trend analysis and load forecasting. Equipment fault records and alarm data include remote diagnostic and alarm modules for all key equipment. When a fault occurs, alarm information is sent in real time, and the fault type, occurrence time, and handling status are recorded. This information is stored in a historical database for subsequent analysis. Data transmission and centralized storage: Each sensor and monitoring device uploads real-time data to a data acquisition server via Ethernet or wireless network. This server uses standard communication protocols (such as Modbus, OPC UA, etc.) to achieve unified data acquisition. The data acquisition server stores the received real-time data in a local data warehouse and simultaneously transmits the data to the central data management platform through a secure encrypted channel, providing reliable data support for subsequent data normalization, preprocessing, and load forecasting.

[0054] S2: Preprocessing the collected data includes data normalization and time-series data reconstruction to form standardized time-series data. In some embodiments, the data center obtains electrical parameters (such as high-voltage system voltage, low-voltage system current, etc.) and environmental parameters (such as computer room temperature, humidity, etc.) of each subsystem through the data acquisition module. It also collects equipment operation logs, load curves, and abnormal alarm data. Due to differences in sensor range, sampling frequency, and data transmission delay, the original data suffers from inconsistent numerical ranges, uneven time intervals, and noise interference. To form standardized time-series data, the preprocessing steps are divided into two parts: data normalization and time-series data reconstruction. First, outlier detection is performed on the collected data. If temperature data exceeds a reasonable range (e.g., above 40 degrees Celsius or below 10 degrees Celsius), these data are considered outliers and are removed or replaced with reasonable estimates. Next, for each type of data, a minimum and a maximum value are determined based on its historical records. Then, each data point is scaled so that all data are mapped to a uniform numerical range, such as from 0 to 1. The purpose of this is to eliminate the influence caused by differences in units and magnitudes between different data items, ensuring that each parameter has equal weight in subsequent processing. Since the collected data may have uneven sampling times or missing data for certain periods, it is necessary to sort the data according to the timestamps attached to each data record. A fixed sampling period (e.g., sampling once per minute) is determined, and the data is resampled. If data is not collected at certain fixed times, interpolation methods (such as spline interpolation) are used to estimate the missing values. For example, if the system records data at 10:00, 10:03, and 10:04, the system will use the data trend between 10:00 and 10:03 to estimate the values ​​at 10:01 and 10:02. Finally, to reduce noise in the data, moving average filtering is used to smooth the reconstructed data, making the data curves more continuous and stable. After the above data normalization and time series data reconstruction processes, the original collected data is transformed into standardized time series data with a fixed sampling period, consistent numerical range, and smooth continuity. These data serve as inputs for subsequent neural network model feature extraction and load forecasting, ensuring data quality and consistency, thereby improving the accuracy and stability of overall load forecasting.

[0055] S3: Feature extraction and load forecasting are performed on the standardized time-series data using a combined neural network model, wherein the combined neural network model includes:

[0056] CNN module: includes at least one convolutional layer and one pooling layer, wherein the convolutional layer uses convolutional kernels to extract local features from the normalized temporal data;

[0057] LSTM module: includes at least one improved long short-term memory network (LSTM) layer, which receives local features output by the CNN module and captures long-term temporal dependencies of the data, wherein the improved LSTM module includes gating units and residual connections to enhance memory performance;

[0058] Attention mechanism module: used to weight the temporal features output by the LSTM module;

[0059] In some embodiments, the data center has obtained standardized time-series data through the aforementioned data acquisition and preprocessing steps. This data is a sequence of electrical and environmental parameters with a fixed sampling period (e.g., once per minute). Next, a combined neural network model is used to extract features and predict load from this data. The specific implementation process is as follows:

[0060] The CNN module is implemented as follows: First, standardized time-series data is input into the CNN module. This module includes at least one convolutional layer and one pooling layer. In the convolutional layer, multiple 3×3 kernels of fixed size are used to perform sliding convolution operations on the input data to extract local features. Specifically, the convolutional layer can capture local fluctuations in the data within a short time window, such as the difference between the maximum and minimum values ​​of the data over several consecutive minutes, the amplitude of local data fluctuations, and local frequency components. After the convolution operation, a set of local feature maps is obtained. Subsequently, a pooling layer (e.g., max pooling) is used to downsample the feature maps output by the convolutional layer, reducing the data dimensionality and retaining the most significant local feature information. For example, for time-series data with 1440 data points per day (one data point per minute), the feature maps obtained after convolution and pooling can reflect the local changing trends and abnormal fluctuations of electrical parameters over a short period of time.

[0061] The improved LSTM module is implemented as follows. Next, the local feature maps output by the CNN module are reorganized into a format suitable for sequence input and fed into the LSTM module. This LSTM module contains at least one layer of improved Long Short-Term Memory (LSTM) network units, whose main function is to capture long-term temporal dependencies in the data. For example, in electrical load data, some changes may persist for a long time; LSTM can extract these long-term dependency information. To enhance the model's memory performance and alleviate the gradient vanishing problem that may occur in traditional LSTM, this module introduces a residual connection structure. Specifically, in each LSTM unit, an output vector is first obtained through standard gating computation, candidate memory state update, and hidden state computation. Then, this output is added element-wise to the current input vector through a linear mapping (or directly using an identity mapping when the input and output dimensions are the same) to obtain the final output. This residual connection not only helps gradients propagate smoothly in deep networks but also allows low-level feature information to be directly transmitted to higher levels, further improving the ability to capture long-term dependent features.

[0062] The attention mechanism module is implemented as follows: the LSTM module outputs a series of temporal feature vectors, each corresponding to a time step. To improve the sensitivity of the prediction model to key information, the attention mechanism module performs weighted processing on these temporal features. Specifically, the LSTM output at each time step is passed through a fully connected layer to generate a preliminary score, which represents the importance of the feature at that moment to load prediction. The preliminary scores of all time steps are then input into a softmax normalization layer to convert them into normalized attention weights, so that the sum of the weights at each time step is 1.

[0063] Finally, the LSTM output features at each time step are multiplied element-wise with the corresponding attention weights, and all weighted feature vectors are summed to obtain a comprehensive feature vector, which serves as the final feature input for load forecasting. This dynamic weighting method enables the model to automatically identify and highlight those moments that have a significant impact on the prediction results, improving overall prediction accuracy.

[0064] Finally, the load forecast result is output. The comprehensive feature vector, weighted by the attention mechanism, is input into one or more fully connected layers, and the final electrical load forecast value is generated through appropriate activation function mapping. This forecast value can reflect the load change trend over a future period (e.g., the next 15 minutes or 1 hour), providing a basis for real-time control of the data center electrical system. Through the above steps, this embodiment realizes feature extraction and load forecasting of standardized time-series data using a combined neural network model. By employing a CNN module to extract local features, an improved LSTM module to capture long-term dependencies (and enhancing memory effects through residual connections), and an attention mechanism module to weight key moment features, the accuracy, stability, and real-time response capability of the forecast are significantly improved.

[0065] The CNN module's specific structure includes the following layers: Input layer: Receives preprocessed, standardized time-series data, typically a one-dimensional or two-dimensional matrix with a fixed sampling period. Convolutional layer: Performs local sliding convolution operations on the input data using multiple fixed-size convolutional kernels (e.g., 3×3 kernels). The number of kernels, stride, and padding can be set according to the specific application requirements, aiming to extract subtle features within a local time window. Activation layer: After convolution, a non-linear activation function (e.g., ReLU) is usually applied to increase the model's non-linear expressive power. Pooling layer: Downsamples the feature map output by the convolutional layer using max pooling or average pooling, reducing data dimensionality and computational cost while preserving key local features. Output layer: Uses the dimensionality-reduced feature map as input to subsequent modules (e.g., LSTM). The CNN module performs local convolution operations by sliding convolutional kernels across the input data, converting local variation information (such as local fluctuations, abrupt changes, frequency components, and statistical features) in the original time-series data into multi-channel feature maps. The introduction of activation functions enables the network to capture non-linear features, while pooling operations further compress information and highlight salient features. Overall, the CNN module plays a role in noise reduction, feature extraction, and dimensionality compression, providing compact feature representations with local semantics for subsequent temporal models.

[0066] In some embodiments, the activation layer of the CNN module employs the Temporally Adaptive Swish Activation Function (TASwish). The traditional Swish activation function is calculated by multiplying the input by its sigmoid transformation, i.e., f(x) = x × sigmoid(β·x), where β is a learnable parameter. To better adapt to the temporal data characteristics in data center electrical load forecasting, this invention introduces a temporal context factor T on the basis of the Swish activation function, thereby obtaining an improved temporally adaptive Swish activation function. Its output calculation method is described as follows: First, the input is set to x, and the output of the activation function is f(x); a lightweight convolutional submodule or attention network is used to extract statistical features (e.g., local mean, variance, maximum, minimum, etc.) from the current local temporal window, and the temporal context factor T is obtained through mapping operations. This factor can reflect the dynamic characteristics of the data at the current moment.

[0067] Introduce T into the activation function and design the formula as follows:

[0068] f(x)=x×sigmoid(β·x+γ·T)

[0069] Here, β and γ are the first and second learnable parameters, which are dynamically adjusted through training. γ is used to control the influence of the time factor T on the output of the activation function.

[0070] Temporal factor extraction: During CNN module processing, after the input standardized temporal data passes through convolutional layers to obtain local feature maps, a lightweight submodule (e.g., 1×1 convolution or a small attention network) performs statistical analysis on these local features to generate a temporal factor T that reflects the dynamic characteristics of the data within the current window. This factor can capture the trends, fluctuations, and anomalies of the current local data.

[0071] Dynamic nonlinear adjustment: By combining the time factor T with the input x, the activation function can dynamically adjust the slope of the nonlinear mapping based on the current data in different time windows. Specifically, when local data experiences significant fluctuations or anomalies, the value of the time factor T changes accordingly, thereby adjusting the input to the sigmoid function and causing the activation function to activate the features at that moment more strongly or less strongly. This dynamic adjustment enhances the sensitivity of the CNN module in capturing subtle changes and sudden features in the data.

[0072] In some embodiments, a specific method for calculating the temporal context factor T is given below. This method extracts statistical features within a local temporal window and then obtains T through a lightweight neural network. The specific steps are as follows:

[0073] Local statistical feature extraction in CNN modules involves calculating multiple statistical metrics within a specific temporal window (e.g., input data over L consecutive time steps or feature maps output by convolutional layers). Commonly used metrics include: Local mean (reflecting the central tendency of the data within the window); Local standard deviation (measuring the volatility or dispersion of the data within the window); and Local maximum and minimum values ​​(capturing extreme values ​​within the window, reflecting anomalies or abrupt changes). These statistical metrics are then concatenated in a fixed order to form a statistical feature vector S, for example, S = [mean, standard deviation, maximum, minimum].

[0074] Lightweight neural network mapping takes a statistical feature vector S as input to a lightweight fully connected network (or MLP), which typically contains only one or two hidden layers. Its purpose is to map multidimensional statistical features to a scalar value, namely the time sequence context factor T.

[0075] Specifically, the following processing flow can be adopted: First layer mapping: Use a fully connected layer to transform S into an intermediate representation vector and connect it to a non-linear activation function (such as ReLU) to enhance the mapping capability. Second layer mapping (optional): Pass it through another fully connected layer to map the intermediate representation into a single scalar. The entire mapping process can be represented as T = f(S), where f(·) represents the composite function of the above fully connected network, and the weights and biases in the network are learnable parameters. After integrating it into the activation function to obtain the temporal context factor T, it is introduced into the improved Swish activation function. The specific implementation is as follows:

[0076] f(x)=x×sigmoid(β·x+γ·T)

[0077] Where x is the input or feature value of the current layer in the CNN module, and β and γ are learnable parameters that are automatically adjusted through training. T plays a role in dynamically adjusting the response of the activation function, enabling the activation function to adaptively change the slope and shape of the nonlinear mapping based on the statistical characteristics of the current local data.

[0078] Throughout the training process of the network, statistical feature extraction and the calculation of T are both part of the forward propagation. During backpropagation, the network automatically updates the parameters of β and γ in the fully connected layers and activation functions based on the prediction error, thereby optimizing the calculation of T and making it better reflect the local temporal features.

[0079] The improved TASwish activation function not only retains the smoothness and non-saturation characteristics of the Swish activation function but also introduces temporal context information, enabling the CNN module to extract local features more accurately when processing electrical load time-series data. After processing by multiple CNN modules, it can more fully extract information such as local fluctuations and abrupt changes, providing richer input features for subsequent LSTM modules and attention mechanisms, thereby improving the overall accuracy and robustness of the load prediction model. By introducing the temporal context factor T, the activation function can adaptively adjust its nonlinear response, dynamically responding to feature changes in data over different time periods. The improved activation function can better capture subtle changes and anomalies in local features, thus improving the feature extraction capability of the CNN module.

[0080] By integrating the entire deep neural network architecture (including the improved LSTM and attention mechanism modules), the application of the TASwish activation function further enhances the accuracy and real-time response capability of data center electrical system load forecasting, providing more accurate data support for system regulation. Through these improvements, the solution introduces an innovative activation function design based on the traditional CNN module. This not only solves the potential expression limitations of ordinary activation functions when processing time-series data but also brings higher prediction accuracy and adaptability to the overall system, further enhancing the innovation and practical value of the technical solution.

[0081] The improved LSTM module's specific structure includes the following aspects: Input layer: Receives local feature maps from the CNN module and reorganizes them into a format suitable for sequence processing according to temporal requirements. LSTM layer: Includes at least one standard LSTM unit, each containing an input gate, a forget gate, and an output gate to control information storage, forgetting, and output. Residual connection: In each LSTM unit, the unit's input (after necessary linear mapping to match the dimension) is element-wise added to its output (hidden state) to form a residual output. This residual connection not only allows low-level information to be directly passed to higher levels but also effectively alleviates the gradient vanishing problem. Output layer: Outputs the temporal features processed by the residual connections and passes them to the attention mechanism module or subsequent fully connected layers. The main function of the LSTM module is to capture long-term dependent information in temporal data. Its internal gating mechanism enables the network to selectively remember or forget information, adapting to the dynamic changes in information within the sequence. The improvement lies in the introduction of residual connections. By directly adding the input to the output of the LSTM unit, the gradient can be propagated more effectively during backpropagation, ensuring the stability and convergence speed of the deep network during training. This structure not only enhances the ability to capture long-term dependent features but also helps maintain the network's sensitivity to low-level features, further improving the overall prediction performance.

[0082] The attention mechanism module's specific structure includes: a fully connected layer that processes the feature vector output by the LSTM module at each time step, generating a preliminary score representing the importance of the current time step's features to the final load prediction; a Softmax normalization layer that inputs the preliminary scores of all time steps into a Softmax function to obtain a normalized attention weight vector, where the sum of these weights is 1, reflecting the relative contribution of each time step to the overall sequence; and a weighted summation layer that multiplies the feature vector of each time step element-wise with its corresponding attention weight, then sums all the weighted feature vectors to obtain a comprehensive feature vector, which is used as the final output for prediction. The core idea of ​​the attention mechanism module is to assign different importance weights to different time steps, allowing the model to automatically focus on the moments most influential on load prediction. The fully connected layer first generates the score for each time step, then the Softmax normalization layer converts the score into probabilistic attention weights. Finally, through weighted summation, the information from all time steps is integrated into a feature vector with global semantics. In this way, when making the final load forecast, the model can utilize the weighted feature information and avoid treating all time steps the same, thereby significantly improving the accuracy and robustness of the forecast.

[0083] S4: Output the predicted electrical load value based on the weighted feature vector. In the preceding steps, after processing by the CNN module, the improved LSTM module, and the attention mechanism module, a comprehensive feature vector with attention weighting has been obtained. This vector contains local information and long-term dependency information extracted from the original time-series data. Next, the weighted feature vector is mapped to the predicted electrical load value through the following steps: Formation of the weighted feature vector: In the attention mechanism module, the LSTM output feature vector at each time step is first processed through a fully connected layer to obtain the corresponding preliminary score. Then, these scores are normalized by Softmax to generate a set of normalized attention weights. The feature vector at each time step is multiplied element-wise with its corresponding attention weight, and the weighted features of all time steps are summed to form a comprehensive weighted feature vector. For example, if the weighted feature vector obtained after attention mechanism processing is F, this vector can fully reflect the contribution of key moments in the sequence to load changes. Fully connected layer mapping: The weighted feature vector F is input into one or more fully connected layers. To enhance the nonlinear mapping capability, a nonlinear activation function (such as ReLU) is usually set between the fully connected layers. In this embodiment, a hidden layer is first set up: the weighted feature vector F (assuming its dimension is 128) is input into a fully connected layer with 64 neurons and processed using the ReLU activation function. Then, the output of this hidden layer is passed to the final output layer. The output layer is a fully connected layer with a single neuron, using a linear activation function to generate a continuous predicted value, i.e., the predicted value of the electrical load. The load prediction output, the final single continuous value, is the predicted electrical load. For example, if the output value is 0.75, after inverse normalization, it may correspond to an electrical load of 750 kilowatts at a certain moment. During the model training phase, the system compares the predicted output with the actual load value, calculates the prediction error using loss functions such as mean squared error (MSE), and updates the network parameters using the backpropagation algorithm. The trained model can input the latest standardized time-series data in real time, and through the above fully connected layer, output instant prediction results, providing a basis for real-time control of the data center electrical system. Once the model outputs the predicted load value, this value can be transmitted to the data center's control system, which can then automatically adjust power allocation, activate backup equipment, or optimize energy management based on the prediction results. Simultaneously, the system continuously monitors the error between the actual load and the predicted load, and adaptively updates the model through an online learning mechanism to ensure prediction accuracy and stability during long-term operation.

[0084] In some embodiments, the implementation of outputting the electrical load prediction value based on the weighted feature vector in step S4 includes the following steps: (a) inputting the weighted feature vector output by the attention mechanism module into at least one fully connected layer, wherein the fully connected layer uses the ReLU activation function to perform nonlinear mapping on the input features to generate an intermediate feature representation; (b) further inputting the intermediate feature representation generated in step (a) into an output fully connected layer with a single neuron, wherein the output layer uses a linear activation function to map the intermediate features into a continuous electrical load prediction value; (c) the continuous electrical load prediction value obtained by the output layer is used as the final prediction result, and after inverse normalization processing, it is used for real-time control of the data center electrical system; (d) during model training, by using the error between the predicted value and the actual load data (e.g., mean square error) as the loss function, the weights of the fully connected layer are adjusted online using the backpropagation algorithm to achieve online updating and adaptive optimization of the prediction model. Through the above steps, the weighted feature vector enhanced by the attention mechanism is mapped into an accurate electrical load prediction value, thereby providing high-precision, real-time load prediction support for the data center electrical system.

[0085] S5: Implement real-time control of the data center electrical system based on the predicted results. The real-time control steps specifically include the following steps: (a) Transmit the predicted electrical load value output by the combined neural network model to the control unit of the data center electrical system. The control unit is interconnected with the main distribution board, UPS equipment, backup power supply and other regulating equipment through a standard communication interface (such as Modbus, OPC UA or Ethernet protocol); (b) After receiving the predicted value, the control unit compares the value with a preset load threshold to determine whether there is a risk of overload or underload in the future; (c) When the predicted value exceeds the preset upper limit threshold, the control unit automatically generates a control command, for example: (i) achieve load balancing by adjusting the transformer tap or redistributing the power lines; (ii) start the backup UPS or backup generator to share the high load risk; (iii) activate the load reduction strategy of some high-power equipment to delay or reduce the load output of some equipment.

[0086] (d) When the predicted value is lower than the preset lower limit threshold, the control unit automatically generates corresponding energy efficiency optimization instructions: (i) shuts down some backup power devices in a timely manner to reduce unnecessary energy consumption; (ii) adjusts the equipment operation mode to enable the electrical system to operate in a more energy-efficient state; (e) the control unit sends the generated control instructions to each relevant device through the real-time data bus and monitors the response status of each device in real time to form a closed-loop control system so as to dynamically optimize and adjust the control strategy based on the feedback information; (f) the system compares and analyzes the actual load data with the predicted load data to form error feedback, which is used by the subsequent online learning module to adaptively update the prediction model parameters.

[0087] In some embodiments, the data center deploys an intelligent power management system and a centralized control unit. This control unit periodically receives electrical load data predicted by a deep neural network model for the next 15 minutes. For example, if the predicted load shows a potential peak of 950 kW within the next 15 minutes, while the system's set safety limit is 900 kW, the control unit immediately issues a control command through its internal control program. This command instructs the main distribution board to reallocate some of the load to the backup power supply, and simultaneously instructs the backup UPS system to be activated to alleviate load pressure. Conversely, when the predicted load value is below a preset lower limit (e.g., below 500 kW), the control unit issues a power reduction command, shutting down some backup equipment to reduce energy consumption. Through real-time communication with each device, the control unit continuously monitors the control effect and adjusts the control strategy based on actual load feedback, thereby ensuring that the data center's electrical system always operates safely, stably, and efficiently.

[0088] In some embodiments, the electrical parameters include: voltage values, current values, power factor, system frequency, harmonic content, active power and reactive power of the high-voltage system and the low-voltage system; the environmental parameters include: temperature, humidity, air velocity and air pressure in the data center computer room; the historical operating data includes: operating logs, load curve data, equipment fault records and abnormal alarm data of each subsystem of the data center.

[0089] In some embodiments, the data normalization includes applying min-max normalization or z-score normalization to the cleaned electrical and environmental parameters to map the data values ​​to a predetermined numerical range; the time-series data reconstruction includes sorting the data according to the acquisition timestamps of each data item, reconstructing the data with inconsistent sampling intervals into time-series data with a fixed sampling period using spline interpolation algorithm, and further reducing data noise by using moving average filtering to form standardized time-series data.

[0090] In some embodiments, after the data acquisition module obtains raw data from various sensors, it performs preliminary data cleaning to remove obvious anomalies and missing data, resulting in a relatively complete dataset. This dataset contains various data such as temperature, humidity, voltage, and current, and these data have different dimensions and numerical ranges. For example, temperature may be between 15 and 35 degrees Celsius, while voltage may be between 200 and 500 volts. To eliminate the dimensional differences between different data, data normalization is performed first. Data normalization first involves determining the historical data range for each cleaned data item. For example, the dynamic range of temperature is determined by statistically analyzing the minimum and maximum values ​​of temperature data over a period of time; similarly, the minimum and maximum values ​​are determined for voltage, current, and other data. Then, a min-max normalization method is used to linearly map each data item according to its minimum and maximum values, so that the value of each data point is mapped to a uniform interval (e.g., 0 to 1). In this way, all data items are on the same scale, facilitating subsequent feature extraction and model training. Alternatively, z-score standardization can be used, which transforms the data into a distribution with a mean of zero and a standard deviation of one by subtracting the mean and then dividing by the standard deviation. Ultimately, regardless of the method used, the goal is to ensure that the values ​​of different data items fall within the same numerical range.

[0091] Because the timestamps of the collected data may be discontinuous or inconsistent (e.g., data is missing at certain times due to network latency or sensor malfunction), all data records are first strictly sorted according to their timestamps. A fixed sampling period is determined, such as sampling once per minute. For time points where data could not be collected at the scheduled time, spline interpolation is used to smooth the missing data. For example, if data was recorded at 10:00, 10:03, and 10:04, the data trend between 10:00 and 10:03 is used to estimate the values ​​at 10:01 and 10:02 using spline interpolation, thus filling the gaps in the time series. The reconstructed time series data may still contain noise fluctuations. To further smooth the data, a moving average filtering technique is used. Specifically, the average of data within a continuous fixed time window (e.g., the past 5 minutes) is used to replace the original value at the current time, making the data curve smoother and more continuous, and reducing the impact of sudden outliers.

[0092] After the data normalization and time-series data reconstruction processes described above, the original multi-source data is transformed into standardized time-series data with a uniform numerical range, fixed sampling period, and low noise interference. This data format retains key feature information and provides high-quality, stable input data for subsequent feature extraction and load prediction in deep neural network models.

[0093] In some embodiments, the convolutional layers in the CNN module employ multiple 3×3 convolutional kernels of a fixed size, extracting local features from temporal data using a sliding window approach, and the pooling layers employ max pooling to reduce data dimensionality and computational complexity. The local features include local fluctuation features, abrupt change features, local frequency components, and local statistical features. The local fluctuation features include the difference between the maximum and minimum values ​​of data within the fixed sliding window and the standard deviation of the data within the fixed sliding window. The abrupt change features include the maximum absolute difference between adjacent sampling points within the fixed sliding window. The local frequency components include the frequency component with the largest amplitude determined by applying a discrete Fourier transform to the data within the fixed sliding window. The local statistical features include the arithmetic mean and standard deviation within the local window.

[0094] In some embodiments, the improved LSTM module includes a residual connection structure, which is specifically implemented as follows:

[0095] First, in each LSTM unit, the input vector is processed through standard LSTM operations to obtain the hidden state output. Standard LSTM operations include gating computation, candidate memory state update, and hidden state computation.

[0096] Secondly, the residual connection structure will convert the input vector x of the current LSTM unit into a single input vector. t With output hidden state h t Perform element-wise addition to form the residual output y. t ;

[0097] y t =h t +f(x t )

[0098] Wherein, the function f(x) t ) is a linear mapping operation used to transform the input vector x t Mapped to h t A feature space with the same dimension; when the input dimension and the hidden state dimension are the same, f(x) t Take the identity mapping directly;

[0099] Finally, the residual connection structure not only connects the input and output within the same time step, but also allows the output of the previous LSTM module to be used as the input of the next LSTM module in a multi-layer LSTM structure. By transferring feature information across layers through residual connections, the ability to capture long-term dependent features is enhanced and the gradient vanishing problem is alleviated.

[0100] In some embodiments, the attention mechanism module includes:

[0101] A fully connected layer is used to generate an initial score from the LSTM output feature vector at each time step;

[0102] A softmax normalization layer converts the initial score into normalized attention weights;

[0103] The weighting operation multiplies the attention weights element-wise with the feature vector output by the LSTM to obtain a weighted feature vector, which is used as the final feature for load prediction.

[0104] The fully connected layer of the attention mechanism module has N neurons. The softmax normalization layer normalizes the N-dimensional preliminary score output by the fully connected layer. The normalization result is used as a weight vector and multiplied element-wise with the N-dimensional feature vector output by the LSTM, thereby achieving weighting of features at each time step.

[0105] This invention also provides a load forecasting system based on data center electrical systems, such as... Figure 2 The system includes the following main hardware components, interconnected via industrial Ethernet and fieldbus, forming a complete closed-loop system for data acquisition, processing, prediction, storage, and real-time control. The data acquisition unit includes sensor modules: electrical parameter acquisition sensors: high-precision voltage sensors, current sensors, power factor sensors, frequency detectors, and harmonic analyzers are installed in the high-voltage and low-voltage power distribution rooms of the data center, respectively; active and reactive power sensors are also installed on the main power distribution cabinet and distribution cabinets. Environmental parameter acquisition sensors: temperature and humidity sensors are installed in the computer room; air velocity and air pressure sensors are installed in the air conditioning system and at the computer room entrances and exits. Historical operation data acquisition interface: each key electrical device (such as UPS, distribution board, generator) is equipped with a status monitoring module and intelligent recording device for collecting equipment operation logs, fault records, alarm information, and load curve data. Data acquisition terminal: all sensor data is collected to the data acquisition terminal via RS485, CAN bus, or wireless transmission modules. This terminal is responsible for preliminary data processing, timestamp synchronization, and packaging the data according to a predetermined format.

[0106] The data processing unit includes edge computing devices, such as edge controllers or industrial PCs located in the data center server room. These devices preprocess the data transmitted from the acquisition terminals using local software, including data cleaning, normalization, and time-series data reconstruction. A data gateway is provided, allowing the edge computing devices to upload the preprocessed, standardized time-series data to the central data processing server via industrial Ethernet or wireless LAN (WLAN). The deep neural network prediction unit includes a prediction server—a high-performance server equipped with GPU acceleration—running a combined neural network model (including a CNN module, an improved LSTM module, and an attention mechanism module). This server receives standardized time-series data from the data processing unit and performs real-time electrical load prediction. An online model update module is also included; the server has a built-in online learning algorithm module that dynamically updates the parameters of the deep neural network model based on the error between the actual load and the predicted value, ensuring prediction accuracy.

[0107] The real-time control unit's control terminal includes one or more industrial-grade PLCs or dedicated controllers. It receives load forecast results from the forecasting server, compares them with preset load thresholds, and generates control commands. The control terminal sends these commands to the main power distribution board, UPS equipment, backup generators, and other automated control devices via Modbus, OPC UA, or other industrial communication protocols, enabling real-time control operations such as power distribution, load balancing, and backup equipment start-up and shutdown. The data storage and management unit includes a high-performance data storage server or database system for storing collected raw data, preprocessed data, historical operation records, load forecast results, and system control logs, facilitating subsequent data analysis and model optimization. Data backup and security equipment includes redundant storage and network security devices to ensure data integrity and security. The communication network and human-machine interface connect all hardware modules via industrial Ethernet, forming a high-speed, stable local area network for real-time data transmission and centralized management. For sensors located in areas difficult to wire, wireless communication modules can access the system via wireless gateways. The SCADA system / HMI is equipped with a monitoring workstation and a touch screen display terminal, which displays the collected data, prediction results and control status in real time, allowing operators to monitor the system's operation and make necessary manual interventions.

[0108] Each sensor is connected to the data acquisition terminal via fieldbus or wireless means. The data acquisition terminal communicates with the edge computing device via industrial Ethernet, and the data is uploaded to the central forecasting server after preprocessing. The forecasting server is connected to the real-time control unit via a secure industrial network, and the forecast results are transmitted to the control terminal in real time. The control terminal communicates bidirectionally with each control device (such as UPS and power distribution board) via standard industrial communication protocols to ensure that control commands can be executed quickly and accurately. The data storage server shares data with all system units via a local area network and displays the overall system status to operators through the SCADA system. Through the above hardware configuration and interconnection methods, the entire system realizes a fully automated closed loop from data acquisition, data preprocessing, load forecasting, control execution to data storage and monitoring, providing high-precision, real-time load forecasting and intelligent control support for data center electrical systems.

[0109] The data acquisition module collects electrical parameters, environmental parameters, and historical operating data of each subsystem within the data center.

[0110] The preprocessing module preprocesses the collected data, including data normalization and time-series data reconstruction to form standardized time-series data.

[0111] A combined neural network model computation module performs feature extraction and load prediction on the standardized time-series data. The combined neural network model includes:

[0112] CNN module: includes at least one convolutional layer and one pooling layer, wherein the convolutional layer uses convolutional kernels to extract local features from the normalized temporal data;

[0113] LSTM module: includes at least one improved long short-term memory network (LSTM) layer, which receives local features output by the CNN module and captures long-term temporal dependencies of the data, wherein the improved LSTM module includes gating units and residual connections to enhance memory performance;

[0114] Attention mechanism module: used to weight the temporal features output by the LSTM module;

[0115] The prediction output module outputs the electrical load prediction value based on the weighted feature vector;

[0116] The control module implements real-time control over the data center electrical system based on the predicted results.

[0117] In some embodiments, the electrical parameters include: voltage values, current values, power factor, system frequency, harmonic content, active power and reactive power of the high-voltage system and the low-voltage system; the environmental parameters include: temperature, humidity, air velocity and air pressure in the data center computer room; the historical operating data includes: operating logs, load curve data, equipment fault records and abnormal alarm data of each subsystem of the data center.

[0118] In some embodiments, the data normalization includes applying min-max normalization or z-score normalization to the cleaned electrical and environmental parameters to map the data values ​​to a predetermined numerical range; the time-series data reconstruction includes sorting the data according to the acquisition timestamps of each data item, reconstructing the data with inconsistent sampling intervals into time-series data with a fixed sampling period using spline interpolation algorithm, and further reducing data noise by using moving average filtering to form standardized time-series data.

[0119] In some embodiments, the convolutional layers in the CNN module employ multiple convolutional kernels of a fixed size of 3×3, extract local features from time-series data using a sliding window approach, and the pooling layers employ max pooling to reduce data dimensionality and computational complexity.

[0120] This invention provides a method and system for load forecasting of data center electrical systems, which can achieve the following beneficial technical effects:

[0121] 1. This invention achieves high-precision, real-time prediction of data center electrical system load by integrating technologies such as multi-source data acquisition, data normalization and time-series reconstruction, and combined neural network models. The prediction accuracy is high. It uses CNN to extract local fluctuations, mutations, frequencies and statistical features, combines an improved LSTM to capture long-term time-series dependencies, and further uses an attention mechanism for dynamic weighting. This invention can fully explore the deep-seated patterns in the data, effectively reduce noise interference, and improve the accuracy of load prediction.

[0122] 2. This invention makes full use of multidimensional data, comprehensively reflecting the operating status of the data center by collecting electrical parameters, environmental parameters and historical operating data, providing rich input for the prediction model, ensuring the comprehensiveness and accuracy of data processing and feature extraction, providing decision-making basis for accurate load forecasting and real-time control of the electrical system, helping to optimize power distribution, reduce equipment failure risk, further improve the overall operating efficiency and safety of the data center, and reduce operating costs.

[0123] 3. The improved LSTM module in this invention employs a residual connection structure. This structure effectively alleviates the gradient vanishing problem in deep networks by adding the input vector to the output of the LSTM unit element-wise, ensuring that the gradient can be fully propagated across multiple time steps and layers, thereby significantly enhancing the model's ability to capture long-term dependent features. This improvement not only enhances the training stability and convergence speed of the load prediction model in deep structures, but also makes the model more accurate and robust when facing long-term complex time-series data from data center electrical systems in practical applications. Furthermore, through residual connections, the model can better integrate low-level features and high-level semantic information, providing richer and more stable feature inputs for subsequent attention mechanism modules, thereby further improving the overall system's real-time prediction performance and adaptive control capabilities. In summary, the improved LSTM module and its residual connection structure significantly improve the limitations of traditional LSTM in deep neural networks, enabling this invention to achieve higher accuracy, stability, and real-time response capabilities in data center electrical system load prediction, effectively ensuring the safety and efficiency of data center operations.

[0124] The above provides a detailed description of a load forecasting method and system based on data center electrical systems. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas and methods of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A load forecasting method for data center electrical systems, characterized in that, Including the following steps: S1: The data acquisition module collects electrical parameters, environmental parameters, and historical operating data of each subsystem within the data center; S2: Preprocess the collected data, including data normalization and time series data reconstruction to form standardized time series data; S3: Feature extraction and load forecasting are performed on the standardized time-series data using a combined neural network model, wherein the combined neural network model includes: CNN module: includes at least one convolutional layer and one pooling layer, wherein the convolutional layer uses convolutional kernels to extract local features from the normalized temporal data; LSTM module: includes at least one improved long short-term memory network (LSTM) layer, which receives local features output by the CNN module and captures long-term temporal dependencies of the data, wherein the improved LSTM module includes gating units and residual connections to enhance memory performance; Attention mechanism module: used to weight the temporal features output by the LSTM module; S4: Output the predicted electrical load value based on the weighted feature vector; S5: Implement real-time control of the data center electrical system based on the predicted results.

2. The load forecasting method for data center electrical systems as described in claim 1, characterized in that, The electrical parameters include: voltage, current, power factor, system frequency, harmonic content, active power and reactive power of the high-voltage system and low-voltage system; the environmental parameters include: temperature, humidity, air velocity and air pressure in the data center computer room; the historical operating data includes: operating logs, load curve data, equipment fault records and abnormal alarm data of each subsystem of the data center.

3. The load forecasting method for data center electrical systems as described in claim 1, characterized in that, The data normalization includes mapping the cleaned electrical and environmental parameters to a predetermined numerical range using min-max normalization or z-score normalization methods. The time-series data reconstruction includes sorting the data based on the collection timestamps of each data item, reconstructing data with inconsistent sampling intervals into time-series data with a fixed sampling period using spline interpolation algorithm, and further reducing data noise using moving average filtering to form standardized time-series data.

4. The load forecasting method for data center electrical systems as described in claim 1, characterized in that, The convolutional layers in the CNN module employ multiple 3×3 convolutional kernels of a fixed size. They extract local features from the temporal data using a sliding window approach, and the pooling layers utilize max pooling to reduce data dimensionality and computational complexity. These local features include local fluctuation features, abrupt change features, local frequency components, and local statistical features. Local fluctuation features include the difference between the maximum and minimum values ​​of the data within the fixed sliding window, as well as the standard deviation of the data within the fixed sliding window. Abrupt change features include the maximum absolute difference between adjacent sampling points within the fixed sliding window. Local frequency components include the frequency component with the largest amplitude determined by applying a discrete Fourier transform to the data within the fixed sliding window. Local statistical features include the arithmetic mean and standard deviation within the local window.

5. The load forecasting method for data center electrical systems as described in claim 1, characterized in that, The improved LSTM module includes a residual connection structure, which is specifically implemented as follows: First, in each LSTM unit, the input vector is processed through standard LSTM operations to obtain the hidden state output. Standard LSTM operations include gating computation, candidate memory state update, and hidden state computation. Secondly, the residual connection structure will convert the input vector x of the current LSTM unit into a single input vector. t With output hidden state h t Perform element-wise addition to form the residual output y. t ; y t =h t +f(x t ) Wherein, the function f(x) t ) is a linear mapping operation used to transform the input vector x t Mapped to h t A feature space with the same dimension; when the input dimension and the hidden state dimension are the same, f(x) t Take the identity mapping directly; Finally, the residual connection structure not only connects the input and output within the same time step, but also allows the output of the previous LSTM module to be used as the input of the next LSTM module in a multi-layer LSTM structure. By transferring feature information across layers through residual connections, the ability to capture long-term dependent features is enhanced and the gradient vanishing problem is alleviated.

6. The load forecasting method for data center electrical systems as described in claim 1, characterized in that, The attention mechanism module includes: A fully connected layer is used to generate an initial score from the LSTM output feature vector at each time step; A softmax normalization layer converts the initial score into normalized attention weights; The weighting operation multiplies the attention weights element-wise with the feature vector output by the LSTM to obtain a weighted feature vector, which is used as the final feature for load prediction. The fully connected layer of the attention mechanism module has N neurons. The softmax normalization layer normalizes the N-dimensional preliminary score output by the fully connected layer. The normalization result is used as a weight vector and multiplied element-wise with the N-dimensional feature vector output by the LSTM, thereby achieving weighting of features at each time step.

7. A load forecasting system based on data center electrical systems, characterized in that, include: The data acquisition module collects electrical parameters, environmental parameters, and historical operating data of each subsystem within the data center. The preprocessing module preprocesses the collected data, including data normalization and time-series data reconstruction to form standardized time-series data. A combined neural network model computation module performs feature extraction and load prediction on the standardized time-series data. The combined neural network model includes: CNN module: includes at least one convolutional layer and one pooling layer, wherein the convolutional layer uses convolutional kernels to extract local features from the normalized temporal data; LSTM module: includes at least one improved long short-term memory network (LSTM) layer, which receives local features output by the CNN module and captures long-term temporal dependencies of the data, wherein the improved LSTM module includes gating units and residual connections to enhance memory performance; Attention mechanism module: used to weight the temporal features output by the LSTM module; The prediction output module outputs the electrical load prediction value based on the weighted feature vector; The control module implements real-time control over the data center electrical system based on the predicted results.

8. A load forecasting system based on a data center electrical system as described in claim 7, characterized in that, The electrical parameters include: voltage, current, power factor, system frequency, harmonic content, active power and reactive power of the high-voltage system and low-voltage system; the environmental parameters include: temperature, humidity, air velocity and air pressure in the data center computer room; the historical operating data includes: operating logs, load curve data, equipment fault records and abnormal alarm data of each subsystem of the data center.

9. A load forecasting system based on a data center electrical system as described in claim 7, characterized in that, The data normalization includes mapping the cleaned electrical and environmental parameters to a predetermined numerical range using min-max normalization or z-score normalization methods. The time-series data reconstruction includes sorting the data based on the collection timestamps of each data item, reconstructing data with inconsistent sampling intervals into time-series data with a fixed sampling period using spline interpolation algorithm, and further reducing data noise using moving average filtering to form standardized time-series data.

10. A load forecasting system based on a data center electrical system as described in claim 7, characterized in that, The convolutional layers in the CNN module use multiple 3×3 convolutional kernels of a fixed size to extract local features from time-series data through a sliding window approach, and the pooling layers use max pooling to reduce data dimensionality and computational complexity.

Images