UPS capacity setting method and system for micro-module computer room based on load prediction

Abstract

The present invention provides a UPS capacity setting method and system for a micro-module computer room based on load prediction, which relates to the technical field of uninterruptible power supply, including constructing a load prediction model, dynamically adjusting the weight coefficient of each sub-model, performing multi-dimensional comprehensive modeling, predicting multi-granularity load changes, and adaptively correcting current prediction results to generate load prediction results; integrating a dynamic game mechanism and a Monte Carlo simulation risk assessment engine, constructing a multi-objective UPS capacity configuration model, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a UPS optimal capacity configuration scheme; applying the UPS optimal capacity configuration scheme to the UPS system configuration of the micro-module computer room, constructing a distributed UPS group that meets the requirements of different load partitions, configuring load sharing control based on active current sharing, and dynamically adjusting the distributed UPS group operation mode and capacity allocation according to the load prediction results and in combination with the real-time operation status.

Description

Technical Field

[0001] The present invention relates to the technical field of uninterruptible power supply, and in particular to a method and system for setting the capacity of a micro-module computer room UPS based on load prediction. Background Art

[0002] With the rapid development of data centers and cloud computing technologies, modular computer rooms are becoming more and more popular as an efficient and flexible data center solution. In modular computer rooms, uninterruptible power supply systems (UPS) are key components to ensure the continuous and stable operation of equipment.

[0003] There are some problems with the traditional UPS capacity setting method. The UPS capacity is usually set according to the peak load or estimated maximum load, resulting in energy waste and low equipment utilization, lack of flexibility, increased energy consumption and operating costs; at the same time, static configuration is difficult to balance reliability and energy efficiency, threatening system stability; existing load forecasting methods often use simple statistical models or empirical formulas, resulting in unreasonable UPS capacity settings, lack of intelligent adjustment, and increased operation and maintenance difficulty and cost; there is a lack of collaborative scheduling mechanism between multiple UPS modules, making it difficult to achieve overall optimal capacity utilization and energy efficiency management.

[0004] In summary, there is an urgent need for an intelligent UPS capacity setting method to improve the energy efficiency, reliability and flexibility of micro-module computer rooms. Through accurate prediction, dynamic capacity adjustment, and multi-module collaborative optimization, the operating efficiency and reliability of the UPS system can be improved and the overall cost of ownership of the micro-module computer room can be reduced. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a method and system for setting the capacity of a UPS in a micro-module computer room based on load prediction, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A method for setting UPS capacity in a micro-module computer room based on load prediction is provided, comprising:

[0008] Obtain historical load data of the micro-module computer room and construct a load prediction model, which includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model. An adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, multi-granularity load changes in future time periods are predicted. An error compensation mechanism is introduced to adaptively correct the current prediction results, and a load prediction result containing load demand data at different time scales and different spatial granularities is generated.

[0009] Based on the load prediction results, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the multi-objective optimization, integrating dynamic game mechanism and Monte Carlo simulation risk assessment engine, constructing a multi-objective UPS capacity configuration model, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a dynamically optimized UPS optimal capacity configuration plan;

[0010] The UPS optimal capacity configuration scheme is applied to the UPS system configuration of the micro-module computer room, a distributed UPS group that meets the requirements of different load partitions is constructed, and load sharing control based on active current sharing is configured. According to the load prediction results and combined with the real-time operating status, the operating mode and capacity allocation of the distributed UPS group are dynamically adjusted.

[0011] In an optional embodiment,

[0012] The historical load data of the micro-module computer room is obtained, and a load prediction model is constructed. The load prediction model includes a regression analysis sub-model, a time series analysis sub-model, and a deep neural network sub-model. An adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, multi-granularity load changes in future time periods are predicted, and an error compensation mechanism is introduced to adaptively correct the current prediction results. The load prediction results generated contain load demand data at different time scales and different spatial granularities, including:

[0013] Based on the historical load data, a first-layer load prediction model is constructed, and the first-layer load prediction model includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model, wherein the regression analysis sub-model adopts a multivariate linear regression algorithm, and determines the linear relationship between the load and the various influencing factors by modeling the various influencing factors that affect the load; the time series analysis sub-model is constructed based on the ARIMA model, and captures the time dynamic characteristics of the load by analyzing the autocorrelation and stationarity of the load time series; the deep neural network sub-model is constructed based on the LSTM network, captures the dependency of long sequence data, and learns the deep-level characteristics and laws of load changes; the outputs of the regression analysis sub-model, the time series analysis sub-model and the deep neural network sub-model are used as the outputs of the first-layer load prediction model to obtain the first-layer load prediction result;

[0014] The first-layer load prediction results are used as input features to construct a second-layer load prediction model including a meta-learner. The meta-learner is constructed using an extreme gradient boosting algorithm, and the complex patterns of the first-layer load prediction results are learned by integrating multiple decision trees. In the training process of the meta-learner, an adaptive weight adjustment mechanism based on an online learning method is integrated, and the adaptive weight adjustment mechanism is used to dynamically update the parameters of the meta-learner according to the real-time prediction error of the meta-learner. The updated meta-learner is used to predict the first-layer load prediction results input, obtain the second-layer load prediction results, and generate the load prediction results at the micro-module computer room level.

[0015] Taking the load prediction result at the micro-module computer room level as input, and using the deep neural network sub-model to decompose the load prediction result at the micro-module computer room level into load prediction results at each rack level;

[0016] Taking the rack-level load prediction result as input, and using the regression analysis sub-model to further decompose the rack-level load prediction result into a device-level load prediction result;

[0017] According to the load prediction result at the micromodule computer room level, the load prediction result at the rack level and the load prediction result at the equipment level, a load prediction result including load demand data at different time scales and different spatial granularities is generated.

[0018] In an optional embodiment,

[0019] The first-layer load prediction results are used as input features to construct a second-layer load prediction model including a meta-learner. The meta-learner is constructed using an extreme gradient boosting algorithm. The complex pattern of the first-layer load prediction results is learned by integrating multiple decision trees, including:

[0020] The first layer of load prediction results are used as input features, and the data set is divided into a training set and a validation set by cross-validation;

[0021] Based on the training set, an adaptive weighted sampling method is used to generate a plurality of training subsets with different distributions, and a decision tree is generated based on each of the training subsets. By dynamically adjusting the sampling weight of each of the training subsets, a plurality of decision trees based on different data distributions are obtained to determine the base learner corresponding to the meta learner;

[0022] Based on each decision tree, a multi-level feature interaction mechanism is constructed to calculate the cross terms and nonlinear transformations of the input features, capture the composite nonlinear relationship in the first-layer load prediction results, and obtain an enhanced expression base learner;

[0023] According to the residual of the fitting loss function in the negative gradient direction, the structure and parameters of the decision tree corresponding to each enhanced expression base learner are iteratively optimized, and the performance of the enhanced expression base learner is evaluated on the validation set, and the growth of each decision tree is controlled by combining the regularization term and the early stopping mechanism to obtain a generalized base learner;

[0024] Adopting an adaptive learning rate strategy, the learning rate of each generalized base learner is dynamically adjusted according to the training progress and performance of the decision tree corresponding to each generalized base learner, and the optimal base learner of each decision tree is determined;

[0025] The optimal base learners are weighted and integrated to determine a meta learner, and the meta learner generates a prediction result of the second-layer load prediction model by learning the first-layer load prediction result.

[0026] In an optional embodiment,

[0027] Based on the load forecast results, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the multi-objective optimization, the dynamic game mechanism and Monte Carlo simulation risk assessment engine are integrated to build a multi-objective UPS capacity configuration model, adaptively adjust the UPS response time and capacity scheduling strategy, and obtain the dynamically optimized UPS optimal capacity configuration solution including:

[0028] The weighting factor of the time dimension is introduced, and the first objective function is determined with the goal of minimizing the average utilization of all UPS at all times; the maximum value of the difference between the actual redundancy and the redundancy requirement threshold is calculated, the maximum difference is determined, and the larger value between the maximum difference and 0 is minimized to determine the second objective function; the third objective function is determined with the goal of minimizing the product of the total capacity of all UPS and the unit capacity energy consumption cost coefficient;

[0029] A dynamic game mechanism is introduced. Based on the first objective function, the second objective function and the third objective function, the corresponding three agents are determined, and the utility function corresponding to each agent is determined. The weight coefficient of the objective function corresponding to each agent is used as the strategy space of each agent. Through the strategy interaction and dynamic adjustment between agents, the Logit dynamic strategy update rule is adopted. Each agent dynamically adjusts its corresponding strategy according to the strategy selection of other agents to maximize its own utility function. Through multiple rounds of iterative games, it converges to the Nash equilibrium point, determines the weight coefficient of the objective function corresponding to each agent, and obtains the initial UPS capacity configuration plan.

[0030] Performing Monte Carlo simulation risk assessment on the initial UPS capacity configuration scheme, performing multiple random sampling on the load forecast results, generating multiple load scenarios, and for each load scenario, calculating the capacity missing amount and missing duration under the initial UPS capacity configuration scheme, obtaining the corresponding empirical distribution, assessing the risk level of the initial UPS capacity configuration scheme, and introducing tolerance thresholds for the capacity missing amount and missing duration and the corresponding risk threshold as risk constraints;

[0031] Based on the multi-objective UPS capacity configuration model, dynamic game mechanism and Monte Carlo simulation risk assessment, the UPS capacity dynamic configuration algorithm is determined, the UPS response time and capacity scheduling strategy are adaptively adjusted, and the dynamically optimized UPS optimal capacity configuration plan is obtained.

[0032] In an optional embodiment,

[0033] The UPS capacity dynamic configuration algorithm includes:

[0034] Initialize the UPS capacity configuration plan and use the dynamic game mechanism to update the weight coefficients of each optimization target; solve the multi-objective UPS capacity configuration model based on the weight coefficients to obtain a new UPS capacity configuration plan;

[0035] Conduct Monte Carlo simulation risk assessment on new UPS capacity configuration solutions, calculate the capacity loss probability based on the preset capacity loss tolerance threshold, and calculate the loss duration probability based on the preset loss duration tolerance threshold;

[0036] Determine whether the probability of capacity loss and the probability of loss duration are less than the risk threshold. If the probability of capacity loss is less than the first risk threshold, and the probability of loss duration is less than the second risk threshold, output the optimal UPS capacity configuration plan; if the probability of capacity loss is greater than or equal to the first risk threshold, or the probability of loss duration is greater than or equal to the second risk threshold, adjust the UPS response time and capacity scheduling strategy, and return to the step of solving the multi-objective UPS capacity configuration model based on the weight coefficient;

[0037] Repeat the iteration until the preset number of iterations is reached to determine the optimal UPS capacity configuration plan.

[0038] In an optional embodiment,

[0039] Active current balancing based load sharing control includes:

[0040] The power sensor equipped in each UPS module collects the input voltage, input current, output voltage, output current and power factor of the UPS module in real time, and transmits them to the central controller via industrial Ethernet;

[0041] Calculate the instantaneous load rate according to the real-time output power and rated capacity of each UPS module, calculate the moving average load rate of each UPS module within a preset time period based on the instantaneous load rate, determine the average value of the moving average load rate, calculate the sum of squares of the difference between the moving average load rate of each UPS module and the average value, determine the arithmetic mean of the sum of squares, calculate the corresponding square root, and determine the load balancing index;

[0042] Comparing the load balance index with a preset threshold, and when the load balance index is greater than the preset threshold, triggering a power reallocation process and determining a power reallocation result;

[0043] Based on the power redistribution result, a power adjustment command including the target power value and the adjustment rate is generated, and the power adjustment command is sent to the control unit of each UPS module through a secure encrypted channel until the control adjustment is completed, and each UPS module sends a confirmation message to the central controller.

[0044] In an optional embodiment,

[0045] The power reallocation process includes:

[0046] The UPS modules are sorted from high to low according to the moving average load rate, and starting from the UPS module with the highest load rate, the amount of power transferred to the UPS module with the lowest load rate is calculated, wherein the power amount is calculated as follows:

[0047] Calculate the power difference between the UPS module with the highest load rate and the UPS module with the lowest load rate; calculate half of the rated capacity of the UPS module with the lowest load rate to determine the half capacity; calculate the product of the power difference and the half capacity; compare the power difference with the product, select the minimum value, and determine the power amount;

[0048] The calculation is repeated and the load rate of each UPS module is updated until the load balance index is less than or equal to the preset threshold, and the power redistribution result is determined.

[0049] According to a second aspect of the embodiments of the present invention,

[0050] A UPS capacity setting system for a micro-module computer room based on load prediction is provided, comprising:

[0051] The first unit is used to obtain the historical load data of the micro-module computer room and build a load prediction model. The load prediction model includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model. The adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, the multi-granularity load changes in the future time period are predicted, and an error compensation mechanism is introduced to adaptively correct the current prediction results to generate a load prediction result containing load demand data at different time scales and different spatial granularities;

[0052] The second unit is used to build a multi-objective UPS capacity configuration model based on the load prediction result, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the optimization multi-objective, integrating dynamic game mechanism and Monte Carlo simulation risk assessment engine, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a dynamically optimized UPS optimal capacity configuration plan;

[0053] The third unit is used to apply the UPS capacity configuration scheme to the UPS system configuration of the micro-module computer room, build a distributed UPS group that meets the needs of different load partitions, and configure load sharing control based on active current sharing. According to the load prediction results and combined with the real-time operating status, the distributed UPS group operating mode and capacity allocation are dynamically adjusted.

[0054] According to a third aspect of the embodiments of the present invention,

[0055] An electronic device is provided, comprising:

[0056] processor;

[0057] a memory for storing processor-executable instructions;

[0058] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0059] According to a fourth aspect of the embodiments of the present invention,

[0060] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0061] In the embodiment of the present invention, by integrating regression analysis, time series analysis and deep neural network sub-models, and adopting an adaptive weight adjustment mechanism, the weight of each sub-model is automatically adjusted according to different scenarios, which can effectively improve the accuracy of load prediction; the multi-sub-model comprehensive modeling method supports flexible prediction of different types of load data and adapts to the complexity and variability of load demand; the method can generate load prediction results at different time scales and spatial granularities to meet the load demand prediction of the computer room in various time ranges such as short-term, medium-term and long-term; by comprehensively considering UPS capacity utilization, redundancy requirements and energy consumption costs, multi-objective optimization is achieved to improve the overall system performance; the integrated dynamic game mechanism can adapt to load in real time Changes can be made to optimize the response time and capacity scheduling strategy of UPS and enhance the flexibility of the system. The Monte Carlo simulation risk assessment engine provides quantitative analysis to help identify and evaluate potential risks in capacity configuration and improve the scientific nature of decision-making. By building a distributed UPS group that meets the needs of different load partitions, flexible load partition management can be achieved to enhance the adaptability of the system. Active current-sharing-based load sharing control can be configured to effectively distribute the load between UPSs, ensure that the system maintains load balance under different operating conditions, and improve overall operating efficiency. According to the load prediction results and real-time operating status, the operating mode and capacity allocation of the distributed UPS group can be dynamically adjusted to ensure that the UPS system can respond to load changes in a timely manner and optimize resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of a flow chart of a method for setting UPS capacity in a micro-module computer room based on load prediction according to an embodiment of the present invention;

[0063] Figure 2 The present invention is a schematic diagram of the structure of a UPS capacity setting system for a micro-module computer room based on load prediction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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.

[0065] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0066] Figure 1FIG. 1 is a flow chart of a method for setting UPS capacity in a micro-module computer room based on load prediction according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] S101. Obtain historical load data of the micro-module computer room and construct a load prediction model, which includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model. An adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, multi-granularity load changes in future time periods are predicted, and an error compensation mechanism is introduced to adaptively correct the current prediction results to generate load prediction results containing load demand data at different time scales and different spatial granularities;

[0068] In this embodiment, by integrating regression analysis, time series analysis and deep neural network sub-models, and adopting an adaptive weight adjustment mechanism, the weight of each sub-model is automatically adjusted according to different scenarios, which can effectively improve the accuracy of load prediction; the multi-sub-model comprehensive modeling method supports flexible prediction of different types of load data, and adapts to the complexity and variability of load demand; this method can generate load prediction results at different time scales and spatial granularity, and meet the load demand prediction of the computer room in various time ranges such as short-term, medium-term and long-term; introduce an error compensation mechanism to adaptively correct the current prediction results, effectively reduce the prediction error, and further improve the real-time and reliability of the prediction; through multi-dimensional comprehensive modeling of computer room equipment, it can comprehensively predict the load changes in different time periods in the future, which is suitable for a variety of load change modes.

[0069] In an optional implementation, historical load data of the micromodule computer room is obtained to construct a load prediction model, which includes a regression analysis sub-model, a time series analysis sub-model, and a deep neural network sub-model. An adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, multi-granularity load changes in future time periods are predicted, and an error compensation mechanism is introduced to adaptively correct the current prediction results. The load prediction results generated include load demand data at different time scales and different spatial granularities, including:

[0070] Based on the historical load data, a first-layer load prediction model is constructed, and the first-layer load prediction model includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model, wherein the regression analysis sub-model adopts a multivariate linear regression algorithm, and determines the linear relationship between the load and the various influencing factors by modeling the various influencing factors that affect the load; the time series analysis sub-model is constructed based on the ARIMA model, and captures the time dynamic characteristics of the load by analyzing the autocorrelation and stationarity of the load time series; the deep neural network sub-model is constructed based on the LSTM network, captures the dependency of long sequence data, and learns the deep-level characteristics and laws of load changes; the outputs of the regression analysis sub-model, the time series analysis sub-model and the deep neural network sub-model are used as the outputs of the first-layer load prediction model to obtain the first-layer load prediction result;

[0071] The first-layer load prediction results are used as input features to construct a second-layer load prediction model including a meta-learner. The meta-learner is constructed using an extreme gradient boosting algorithm, and the complex patterns of the first-layer load prediction results are learned by integrating multiple decision trees. In the training process of the meta-learner, an adaptive weight adjustment mechanism based on an online learning method is integrated, and the adaptive weight adjustment mechanism is used to dynamically update the parameters of the meta-learner according to the real-time prediction error of the meta-learner. The updated meta-learner is used to predict the first-layer load prediction results input, obtain the second-layer load prediction results, and generate the load prediction results at the micro-module computer room level.

[0072] Taking the load prediction result at the micro-module computer room level as input, and using the deep neural network sub-model to decompose the load prediction result at the micro-module computer room level into load prediction results at each rack level;

[0073] Taking the rack-level load prediction result as input, and using the regression analysis sub-model to further decompose the rack-level load prediction result into a device-level load prediction result;

[0074] According to the load prediction result at the micromodule computer room level, the load prediction result at the rack level and the load prediction result at the equipment level, a load prediction result including load demand data at different time scales and different spatial granularities is generated.

[0075] First, obtain the historical load data of the micro-module room, including multi-dimensional data such as the total load of the room, the load of each rack, and the load of a single device. The time span of the historical data is at least 1 year, and the sampling frequency is once every 5 minutes.

[0076] Based on the historical load data obtained, the first-level load forecasting model is constructed. This model contains three sub-models: regression analysis sub-model, time series analysis sub-model and deep neural network sub-model.

[0077] The regression analysis sub-model uses a multivariate linear regression algorithm. The main factors affecting the load, such as time, temperature, humidity, etc., are selected to construct the independent variable set X. The historical load data is used as the dependent variable Y, and the regression coefficient β is estimated by the least squares method to obtain the regression equation Y = Xβ + ε. This equation can be used to predict future loads.

[0078] The time series analysis sub-model is built based on the ARIMA model. First, the historical load data is differentially processed to make it a stationary time series. Then, the order (p, d, q) of the ARIMA model is determined by analyzing the autocorrelation function and partial autocorrelation function. The maximum likelihood estimation method is used to estimate the model parameters and obtain the ARIMA (p, d, q) model. This model can capture the periodicity and trend characteristics of the load.

[0079] The deep neural network sub-model is built on the LSTM network. The historical load data is divided into several sequence samples in chronological order, and each sample contains the load data of the past 24 hours. The LSTM network consists of an input layer, two hidden layers, and an output layer. The number of neurons in the hidden layers is 64 and 32 respectively. The network parameters are trained by the back-propagation algorithm, so that the network can learn the long-term dependencies of load changes.

[0080] The output results of the above three sub-models are weighted averaged to obtain the output result of the first-layer load prediction model. The initial weights can be set equal, that is, each is 1 / 3.

[0081] Next, the first-layer load prediction results are used as input features to build the second-layer load prediction model. The model uses the extreme gradient boosting (XGBoost) algorithm as a meta-learner. The XGBoost model contains 100 decision trees, each with a maximum depth of 6 and a learning rate of 0.1. Through the nonlinear combination of the first-layer prediction results, the XGBoost model can learn more complex load change patterns.

[0082] During the training process of the XGBoost model, an adaptive weight adjustment mechanism based on online learning is introduced. The specific steps are as follows:

[0083] 1) Initialize the XGBoost model parameters θ;

[0084] 2) For each new sample (x, y), use the current model to make a prediction and get

[0085] 3) Calculate the prediction error

[0086] 4) Based on the prediction error e, the model parameters θ are updated using the Adam optimizer;

[0087] 5) Repeat steps 2)-4) until the model converges;

[0088] In this way, the XGBoost model can dynamically adjust parameters based on the real-time prediction error to improve prediction accuracy.

[0089] The trained XGBoost model is used to predict the first-layer load prediction results to obtain the second-layer load prediction results, that is, the load prediction results at the micro-module computer room level. The prediction time granularity is every 5 minutes, and the prediction range is the next 24 hours.

[0090] The load prediction results at the micromodule room level are used as input and decomposed into the load prediction results at the rack level using a deep neural network sub-model. The neural network contains three fully connected layers with 128, 64, and 32 neurons respectively. The input is the total load of the room and the output is the load of each rack. The network parameters are trained by minimizing the mean square error between the predicted value and the actual value.

[0091] The load prediction results at the rack level are used as input and further decomposed into the load prediction results at the device level using the regression analysis sub-model. For each rack, a multivariate linear regression model is established with independent variables such as time and temperature and dependent variables as the load of each device. The regression coefficients are estimated by the least squares method to obtain the load prediction value of each device.

[0092] Finally, based on the prediction results of the above different granularities, a load prediction result package containing multiple time scales (5 minutes, 1 hour, 1 day) and multiple spatial granularities (equipment, rack, computer room) is generated. The prediction results are stored in JSON format for subsequent processing and analysis.

[0093] In order to further improve the prediction accuracy, an error compensation mechanism is introduced. The predicted value is compared with the actual value regularly (such as every hour) to calculate the error. The error series is input into a simple time series model (such as exponential smoothing) to predict future errors. The predicted error value is superimposed on the original prediction result to obtain the corrected prediction value.

[0094] Through the above method, multi-dimensional and multi-scale prediction of the micro-module room load can be achieved, providing strong support for subsequent energy consumption optimization and capacity planning.

[0095] In this embodiment, by constructing a multi-level prediction model, integrating regression analysis, time series analysis and deep neural network, it is possible to more accurately capture the complex patterns of load changes and improve the overall prediction accuracy; introducing an adaptive weight adjustment mechanism in the meta-learner, it is possible to dynamically adjust the model parameters according to the prediction error in real time, so that the prediction results are more adaptable to the changing load demand; through hierarchical prediction, it is gradually decomposed from the computer room level to the rack level and the equipment level, providing load predictions of different spatial granularities to meet the needs of refined load management; combining the time series analysis of the ARIMA model and the deep feature learning of the LSTM, it is possible to accurately identify the temporal dynamic characteristics and long-sequence dependencies of the load, and improve the grasp of the load change law; by using the meta-learner of the extreme gradient boosting algorithm, the complex patterns of the first-layer prediction results are integrated and learned, thereby improving the generalization ability and prediction effect of the load prediction model; this method generates load prediction results containing different time scales, supporting accurate prediction and planning of short-term and long-term load demands.

[0096] In an optional implementation, the first-layer load prediction results are used as input features to construct a second-layer load prediction model including a meta-learner, and the meta-learner is constructed using an extreme gradient boosting algorithm. The complex pattern of the first-layer load prediction results is learned by integrating multiple decision trees, including:

[0097] The first layer of load prediction results are used as input features, and the data set is divided into a training set and a validation set by cross-validation;

[0098] Based on the training set, an adaptive weighted sampling method is used to generate a plurality of training subsets with different distributions, and a decision tree is generated based on each of the training subsets. By dynamically adjusting the sampling weight of each of the training subsets, a plurality of decision trees based on different data distributions are obtained to determine the base learner corresponding to the meta learner;

[0099] Based on each decision tree, a multi-level feature interaction mechanism is constructed to calculate the cross terms and nonlinear transformations of the input features, capture the composite nonlinear relationship in the first-layer load prediction results, and obtain an enhanced expression base learner;

[0100] According to the residual of the fitting loss function in the negative gradient direction, the structure and parameters of the decision tree corresponding to each enhanced expression base learner are iteratively optimized, and the performance of the enhanced expression base learner is evaluated on the validation set, and the growth of each decision tree is controlled by combining the regularization term and the early stopping mechanism to obtain a generalized base learner;

[0101] Adopting an adaptive learning rate strategy, the learning rate of each generalized base learner is dynamically adjusted according to the training progress and performance of the decision tree corresponding to each generalized base learner, and the optimal base learner of each decision tree is determined;

[0102] The optimal base learners are weighted and integrated to determine a meta learner, and the meta learner generates a prediction result of the second-layer load prediction model by learning the first-layer load prediction result.

[0103] First, the first-layer load prediction results are used as input features, and the data set is divided into a training set and a validation set using a 5-fold cross-validation method. For example, for a data set containing 10,000 samples, it is randomly divided into 5 subsets of equal size, and 4 subsets (8,000 samples) are selected as training sets each time, and the remaining 1 subset (2,000 samples) is used as the validation set, and training and validation are performed alternately for 5 times.

[0104] Next, based on the training set, an adaptive weighted sampling method is used to generate multiple training subsets with different distributions. Specifically, each sample in the training set is first given an initial weight of 1 / N (N is the total number of samples). Then the weight of each sample is adjusted according to its prediction error. The sample with a larger prediction error has a higher weight. A roulette wheel method is used to extract samples from the training set with replacement to form a new training subset. The number of samples in each training subset is 80% of the original training set. In this way, 100 training subsets are generated, each with a different data distribution.

[0105] A decision tree is generated based on each training subset as a base learner. The decision tree is generated using the CART algorithm, which selects the optimal partitioning features and partitioning points by minimizing the Gini coefficient. The maximum depth of the tree is set to 6, and the minimum number of leaf node samples is set to 20. During the training process, the sampling weight of each training subset is dynamically adjusted so that samples that are difficult to classify have a higher probability of being selected, thereby obtaining 100 decision trees based on different data distributions, which constitute the base learner set of the meta-learner.

[0106] For each decision tree, a multi-level feature interaction mechanism is constructed to calculate the cross terms and nonlinear transformations of the input features to capture the complex nonlinear relationship in the first-level load prediction results. Specifically, for each internal node, in addition to the original features, the second-order interaction terms (such as the product of feature A and feature B) and the third-order interaction terms (such as the product of features A, B, and C) between the features are calculated. At the same time, nonlinear transformations are applied to each feature, including logarithmic transformations, exponential transformations, and polynomial transformations. This can greatly enhance the decision tree's ability to express complex patterns and obtain an enhanced expression base learner.

[0107] Then, according to the residual of the fitting loss function in the negative gradient direction, the structure and parameters of the decision tree corresponding to each enhanced expression base learner are iteratively optimized. Specifically, the negative gradient value of each sample is first calculated, and then these negative gradient values ​​are used as new target variables to refit the decision tree. In this way, the structure of the decision tree is gradually optimized so that it can better fit the residual. At the same time, the performance of the enhanced expression base learner is evaluated on the validation set, and the model complexity is controlled by combining L1 and L2 regularization terms, and an early stopping mechanism is used to avoid overfitting. For example, when the performance on the validation set does not improve after 5 consecutive iterations, the optimization of the current decision tree is stopped. In this way, a base learner with good generalization ability is obtained.

[0108] In order to further improve the performance of the model, an adaptive learning rate strategy is adopted to dynamically adjust the learning rate of each generalized base learner according to the training progress and performance of the corresponding decision tree. Specifically, the initial learning rate is set to 0.1, and then the learning rate is adjusted according to the performance changes on the validation set at each iteration. If the performance improves for 3 consecutive iterations, the learning rate is multiplied by 1.1; if the performance decreases for 3 consecutive iterations, the learning rate is multiplied by 0.9. In this way, the optimal base learner for each decision tree is determined.

[0109] Finally, the optimal base learners are weighted and integrated to determine the meta-learner. The integrated weight adopts an adaptive weight allocation method based on the performance of the validation set. The base learners with better performance receive higher weights. The meta-learner generates the prediction results of the second-layer load prediction model by learning the first-layer load prediction results. For example, for a new input sample, 100 base learners are first used to make predictions, and then these 100 prediction results are weighted averaged according to their respective weights to obtain the final prediction value.

[0110] Through the above steps, a powerful second-layer load forecasting model is constructed, which can effectively capture the complex nonlinear patterns in the first-layer forecast results and improve the overall forecasting accuracy. In practical applications, this method has shown excellent performance in various load forecasting tasks, with an average improvement of 10-15% in forecasting accuracy compared to single model methods.

[0111] In this embodiment, multiple training subsets with different distributions are generated through cross-validation and adaptive weighted sampling, thereby enhancing the generalization ability of the model under different data distributions; a multi-level feature interaction mechanism is constructed to calculate the cross terms and nonlinear transformations of the input features, which can effectively capture the complex nonlinear relationship in load forecasting and improve the prediction accuracy; through iterative optimization based on negative gradient fitting residuals, combined with regularization terms and early stopping mechanisms, the structure and parameters of the decision tree are effectively controlled to avoid model overfitting and improve the stability of the model; an adaptive learning rate strategy is adopted to dynamically adjust the learning rate according to the training progress and performance to ensure that the model obtains the best learning effect during the training process; by weighted integration of multiple optimal basis learners, a meta-learner is determined to further improve the prediction accuracy and robustness of the second-layer load forecasting model.

[0112] S102. Based on the load forecast results, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the multi-objective optimization, integrating dynamic game mechanism and Monte Carlo simulation risk assessment engine, constructing a multi-objective UPS capacity configuration model, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a dynamically optimized UPS optimal capacity configuration plan;

[0113] In this embodiment, multi-objective optimization is achieved by comprehensively considering UPS capacity utilization, redundancy requirements and energy consumption costs, and the overall system performance is improved; the integrated dynamic game mechanism can adapt to load changes in real time, optimize the response time and capacity scheduling strategy of UPS, and enhance system flexibility; the Monte Carlo simulation risk assessment engine provides quantitative analysis to help identify and evaluate potential risks in capacity configuration and improve the scientific nature of decision-making; the dynamically optimized UPS optimal capacity configuration plan finally generated improves energy efficiency and cost-effectiveness, and provides effective support for computer room management.

[0114] In an optional implementation, based on the load forecast results, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the multi-objectives to be optimized, integrating dynamic game mechanism and Monte Carlo simulation risk assessment engine, constructing a multi-objective UPS capacity configuration model, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a dynamically optimized UPS optimal capacity configuration solution including:

[0115] The weighting factor of the time dimension is introduced, and the first objective function is determined with the goal of minimizing the average utilization of all UPS at all times; the maximum value of the difference between the actual redundancy and the redundancy requirement threshold is calculated, the maximum difference is determined, and the larger value between the maximum difference and 0 is minimized to determine the second objective function; the third objective function is determined with the goal of minimizing the product of the total capacity of all UPS and the unit capacity energy consumption cost coefficient;

[0116] A dynamic game mechanism is introduced. Based on the first objective function, the second objective function and the third objective function, the corresponding three agents are determined, and the utility function corresponding to each agent is determined. The weight coefficient of the objective function corresponding to each agent is used as the strategy space of each agent. Through the strategy interaction and dynamic adjustment between agents, the Logit dynamic strategy update rule is adopted. Each agent dynamically adjusts its corresponding strategy according to the strategy selection of other agents to maximize its own utility function. Through multiple rounds of iterative games, it converges to the Nash equilibrium point, determines the weight coefficient of the objective function corresponding to each agent, and obtains the initial UPS capacity configuration plan.

[0117] Performing Monte Carlo simulation risk assessment on the initial UPS capacity configuration scheme, performing multiple random sampling on the load forecast results, generating multiple load scenarios, and for each load scenario, calculating the capacity missing amount and missing duration under the initial UPS capacity configuration scheme, obtaining the corresponding empirical distribution, assessing the risk level of the initial UPS capacity configuration scheme, and introducing tolerance thresholds for the capacity missing amount and missing duration and the corresponding risk threshold as risk constraints;

[0118] Based on the multi-objective UPS capacity configuration model, dynamic game mechanism and Monte Carlo simulation risk assessment, the UPS capacity dynamic configuration algorithm is determined, the UPS response time and capacity scheduling strategy are adaptively adjusted, and the dynamically optimized UPS optimal capacity configuration plan is obtained.

[0119] First, introduce the weighting factor of the time dimension, calculate the average utilization of all UPS at all times, and minimize it as the first objective function. For example, assuming there are 3 UPSs, considering a 24-hour period, the hourly utilization data are as follows: UPS1: [0.6, 0.7, 0.5, ...], UPS2: [0.5, 0.6, 0.4, ...], UPS3: [0.7, 0.8, 0.6, ...]. The weighting factor can be set to [1, 1, 1.2, 1.5, ...] to indicate the importance of different time periods. The weighted average utilization is calculated to be 0.62, which is used as the optimization target of the first objective function.

[0120] Secondly, the maximum value of the difference between the actual redundancy and the redundancy requirement threshold is calculated, and the larger one between the difference and 0 is taken to minimize the second objective function. For example, assuming that the redundancy requirement threshold is 1.2, and the actual redundancy changes within 24 hours is [1.1, 1.3, 1.2, ...], then the difference sequence is [-0.1, 0.1, 0, ...], and the maximum difference is 0.1. The optimization target of the second objective function is min(max(0.1, 0)) = 0.1.

[0121] Again, calculate the product of the total capacity of all UPS and the unit capacity energy consumption cost coefficient, and minimize it as the third objective function. For example, the capacities of the three UPS are 100kVA, 150kVA, and 200kVA, and the unit capacity energy consumption cost coefficient is 0.01, then the optimization target of the third objective function is (100+150+200)*0.01=4.5.

[0122] Next, a dynamic game mechanism is introduced to determine the corresponding three agents based on the above three objective functions. The strategy space of each agent is the weight coefficient of the corresponding objective function. Using the Logit dynamic strategy update rule, each agent dynamically adjusts its own strategy according to the strategy selection of other agents to maximize its own utility function. Through multiple rounds of iterative games, it finally converges to the Nash equilibrium point and determines the weight coefficient of each objective function. For example, after 100 rounds of iterations, the weight coefficients of the three objective functions converge to 0.4, 0.3, and 0.3, respectively, and the initial UPS capacity configuration plan is obtained accordingly.

[0123] Then, a Monte Carlo simulation risk assessment is performed on the initial UPS capacity configuration scheme. The load forecast results are randomly sampled 1,000 times to generate 1,000 load scenarios. For each scenario, the capacity missing amount and missing duration under the initial configuration scheme are calculated to obtain the corresponding empirical distribution. For example, the 95% quantile of the capacity missing amount is 50kVA, and the 90% quantile of the missing duration is 30 minutes. The capacity missing amount tolerance threshold of 100kVA and the missing duration tolerance threshold of 60 minutes, as well as the corresponding risk thresholds of 5% and 10%, are introduced as risk constraints.

[0124] Finally, based on the multi-objective UPS capacity configuration model, dynamic game mechanism and Monte Carlo simulation risk assessment, the UPS capacity dynamic configuration algorithm is determined. The algorithm minimizes the weighted sum of the three objective functions by continuously adjusting the UPS response time and capacity scheduling strategy, while satisfying the risk constraints. For example, the capacity of UPS1 is adjusted from 100kVA to 120kVA, the capacity of UPS2 is adjusted from 150kVA to 140kVA, the capacity of UPS3 remains unchanged at 200kVA, and the response time of UPS is adjusted accordingly, so that the risk of capacity loss is reduced to an acceptable level. Through multiple rounds of iterative optimization, the dynamically optimized UPS optimal capacity configuration solution is finally obtained.

[0125] This method combines multi-objective optimization, dynamic game and risk assessment to achieve dynamic optimization configuration of UPS capacity. It can improve UPS capacity utilization and reduce energy consumption costs while ensuring power supply reliability, providing effective technical support for the efficient operation of UPS systems in data centers.

[0126] In this embodiment, by introducing three objective functions, comprehensive optimization is performed on average utilization, redundancy difference and energy consumption cost to improve the overall efficiency of the UPS system; by applying a dynamic game mechanism, each intelligent agent continuously optimizes its own strategy in strategy interaction, promotes the system to reach Nash equilibrium in multiple rounds of iterations, and improves the intelligence and flexibility of decision-making; Monte Carlo simulation provides risk assessment of the initial UPS capacity configuration plan, and can perform multiple random samplings for different load scenarios to quantify the capacity loss and the duration of the loss, thereby enhancing the reliability of the decision; by introducing tolerance thresholds for the capacity loss and the duration of the loss as risk constraints, it is ensured that the UPS configuration plan operates within the risk control range; based on the above mechanism, the response time and capacity scheduling strategy of the UPS can be adaptively adjusted, and finally a dynamically optimized UPS optimal capacity configuration plan is obtained, thereby improving the stability and efficiency of the system.

[0127] In an optional implementation, the UPS capacity dynamic configuration algorithm includes:

[0128] Initialize the UPS capacity configuration plan and use the dynamic game mechanism to update the weight coefficients of each optimization target; solve the multi-objective UPS capacity configuration model based on the weight coefficients to obtain a new UPS capacity configuration plan;

[0129] Conduct Monte Carlo simulation risk assessment on new UPS capacity configuration solutions, calculate the capacity loss probability based on the preset capacity loss tolerance threshold, and calculate the loss duration probability based on the preset loss duration tolerance threshold;

[0130] Determine whether the probability of capacity loss and the probability of loss duration are less than the risk threshold. If the probability of capacity loss is less than the first risk threshold, and the probability of loss duration is less than the second risk threshold, output the optimal UPS capacity configuration plan; if the probability of capacity loss is greater than or equal to the first risk threshold, or the probability of loss duration is greater than or equal to the second risk threshold, adjust the UPS response time and capacity scheduling strategy, and return to the step of solving the multi-objective UPS capacity configuration model based on the weight coefficient;

[0131] Repeat the iteration until the preset number of iterations is reached to determine the optimal UPS capacity configuration plan.

[0132] Specifically, the UPS capacity configuration scheme is initialized first, and the initial UPS capacity can be set according to historical data or experience. Then the weight coefficient of each optimization target is updated using a dynamic game mechanism. Specifically, the Nash equilibrium principle in game theory can be used to model the UPS capacity configuration problem as a multi-participant non-cooperative game, where the participants include power supply reliability, economy, environmental protection and other targets. By iteratively calculating the optimal strategy of each participant, the weight coefficient of each target is finally obtained. For example, the reliability weight is 0.5, the economy weight is 0.3, and the environmental protection weight is 0.2.

[0133] Based on the updated weight coefficients, the multi-objective UPS capacity configuration model is solved to obtain a new UPS capacity configuration plan. The multi-objective model includes power supply reliability objective function, economic objective function and environmental objective function. The power supply reliability objective function can be represented by the UPS system availability; the economic objective function can be represented by the total cost of ownership; and the environmental objective function can be represented by carbon emissions. Multiple objective functions are integrated into a single objective function through the weighted summation method, and then solved by intelligent optimization algorithms such as genetic algorithms to obtain a new UPS capacity configuration plan. For example, configure 3 500kVA UPS.

[0134] Perform Monte Carlo simulation risk assessment on new UPS capacity configuration schemes. First, based on historical load data, perform probability modeling of future loads. Then perform a large number of random sampling simulations (e.g., 10,000 times) to simulate the operating status of the UPS system under various load conditions. Based on the preset capacity loss tolerance threshold (e.g., 95% rated capacity), count the number of capacity losses and calculate the capacity loss probability. Similarly, based on the preset loss duration tolerance threshold (e.g., 8 hours per year), count the number of times the threshold is exceeded and calculate the loss duration probability. For example, the capacity loss probability is 0.02 and the loss duration probability is 0.01.

[0135] Determine whether the probability of capacity loss and the probability of missing duration are less than the risk threshold. If the probability of capacity loss is less than the first risk threshold (such as 0.05), and the probability of missing duration is less than the second risk threshold (such as 0.03), the current UPS capacity configuration plan is considered to meet the requirements and can be output as the optimal plan. If the probability of capacity loss is greater than or equal to the first risk threshold, or the probability of missing duration is greater than or equal to the second risk threshold, it is necessary to adjust the UPS response time and capacity scheduling strategy. For example, the UPS startup time can be shortened, or a more flexible UPS modular design can be adopted to improve the dynamic response capability of the system. After adjustment, return to the step of solving the multi-objective UPS capacity configuration model based on the weight coefficient.

[0136] Repeat the above steps until the preset number of iterations (such as 100 times) is reached, or there is no significant improvement after multiple iterations. Finally, the optimal UPS capacity configuration scheme that meets all requirements is determined as the basis for the actual configuration of the system. The entire algorithm process is continuously iterated and optimized to achieve economical and efficient configuration of UPS capacity while ensuring power supply reliability.

[0137] In this embodiment, the weight coefficient of each optimization target is continuously updated through a dynamic game mechanism, and the UPS capacity configuration plan is gradually adjusted to ensure a more accurate optimization process; Monte Carlo simulation risk assessment is introduced, and the capacity loss probability and the loss duration probability are calculated based on the capacity loss tolerance threshold and the loss duration tolerance threshold to ensure that the risk of the UPS capacity configuration plan is within a controllable range; if the assessment result exceeds the risk threshold, the UPS response time and capacity scheduling strategy are adjusted to optimize the system, so that the system has dynamic tuning capabilities to ensure that the UPS configuration adapts to load changes; through multiple rounds of iterative optimization, the configuration plan is gradually adjusted until the best effect is achieved, thereby improving the reliability and efficiency of the UPS capacity configuration; the weight coefficient is used to balance different objectives in the multi-objective optimization model, and finally the optimal UPS capacity configuration plan is output to achieve comprehensive optimization of energy consumption, redundancy and utilization.

[0138] S103. Apply the UPS optimal capacity configuration scheme to the UPS system configuration of the micro-module computer room, build a distributed UPS group that meets the needs of different load partitions, and configure load sharing control based on active current sharing. According to the load prediction results and combined with the real-time operating status, dynamically adjust the distributed UPS group operation mode and capacity allocation.

[0139] First, load zoning planning is carried out according to the physical layout and load characteristics of the micromodule room. The room is divided into critical load area, important load area and ordinary load area. The critical load area contains core network equipment and key business servers, requiring dual-circuit power supply redundancy; the important load area contains important business servers and storage devices, requiring N+1 redundancy; the ordinary load area contains ordinary business equipment and uses a single-circuit power supply.

[0140] Based on the load zoning planning results, a distributed UPS group topology is constructed. Two groups of 20kVA modular UPS are deployed in the key load area, each with 4 UPS modules; a group of 30kVA modular UPS is deployed in the important load area, with 3 UPS modules; a group of 10kVA modular UPS is deployed in the ordinary load area, with 2 UPS modules. All UPS modules use intelligent communication interfaces to support active current sharing control and operation status monitoring.

[0141] Configure active current sharing control strategy to achieve load balancing within the UPS group. Use moving average load rate as the current sharing control quantity, and the control cycle is 1 minute. When the load rate difference between modules exceeds 5%, current sharing adjustment is triggered, and power redistribution is achieved by adjusting the output voltage of each module to ensure that the load rate deviation is controlled within the range of ±2%.

[0142] Establish a distributed UPS group operation status monitoring system. Collect the input voltage, output voltage, output current, power factor, battery status and other operation parameters of each UPS module every second. Calculate the module's real-time load rate, moving average load rate and load forecast deviation, and generate an operation status report.

[0143] According to the short-term load forecast results, the UPS group operation mode is dynamically planned. The forecast period is 24 hours and the forecast granularity is 1 hour. When the predicted load rate is lower than 30%, some UPS modules are switched to standby mode; when the predicted load rate is between 30%-70%, the energy-saving mode is adopted; when the predicted load rate is higher than 70%, it is switched to full power operation mode.

[0144] Monitor load forecast deviation in real time and dynamically adjust UPS group capacity allocation. When the relative deviation between actual load and forecast load exceeds ±10% and lasts for more than 15 minutes, dynamic capacity adjustment is triggered. Fuzzy control algorithm is used to calculate the optimal capacity allocation plan, and smooth switching of UPS modules is achieved through soft start and soft shutdown.

[0145] Optimize the UPS group energy efficiency management strategy. According to the historical operation data, establish the UPS module efficiency curve model. Under the premise of ensuring power supply reliability, adjust the module switch status and load distribution ratio to make each module work in the optimal efficiency range. When the total system load rate is less than 20%, reduce the cumulative operation time of a single module through rotation operation.

[0146] Specifically, the total load of IT equipment in a micro-module computer room is 45kW, including 15kW of critical load, 20kW of important load, and 10kW of ordinary load. Through load forecasting, it is found that during the daytime from 8:00 to 18:00 on weekdays, the total load rate is maintained at 65%-85%; during nighttime and weekends, the total load rate drops to 25%-45%. Based on the forecast results, the energy-saving mode is adopted during the daytime, keeping 2 / 3 of the UPS modules online; at night, it switches to standby mode and only keeps 1 / 2 of the UPS modules online. Through dynamic adjustment, the overall efficiency of the system is improved by 15%, and the annual energy consumption is reduced by about 20%.

[0147] At the same time, an emergency response mechanism is established. When a sudden load change or UPS module failure is detected, the backup module is immediately started to ensure power supply reliability. The system records all operation mode switching and capacity adjustment events, generates operation reports, and provides data support for long-term optimization.

[0148] In this embodiment, by constructing a distributed UPS group that meets the needs of different load partitions, flexible load partition management is achieved and the adaptability of the system is enhanced; load sharing control based on active current sharing is configured to effectively distribute the load between the UPSs, ensuring that the system maintains load balance under different operating conditions and improving the overall operating efficiency; according to the load prediction results and real-time operating status, the operating mode and capacity allocation of the distributed UPS group are dynamically adjusted to ensure that the UPS system can respond to load changes in a timely manner and optimize resource utilization; the configuration of the distributed UPS group enhances the redundancy and fault tolerance of the system, reduces the risk of single point failures, and improves the reliability of the power supply system in the computer room.

[0149] In an optional implementation, the load sharing control based on active current balancing includes:

[0150] The power sensor equipped in each UPS module collects the input voltage, input current, output voltage, output current and power factor of the UPS module in real time, and transmits them to the central controller via industrial Ethernet;

[0151] Calculate the instantaneous load rate according to the real-time output power and rated capacity of each UPS module, calculate the moving average load rate of each UPS module within a preset time period based on the instantaneous load rate, determine the average value of the moving average load rate, calculate the sum of squares of the difference between the moving average load rate of each UPS module and the average value, determine the arithmetic mean of the sum of squares, calculate the corresponding square root, and determine the load balancing index;

[0152] Comparing the load balance index with a preset threshold, and when the load balance index is greater than the preset threshold, triggering a power reallocation process and determining a power reallocation result;

[0153] Based on the power redistribution result, a power adjustment command including the target power value and the adjustment rate is generated, and the power adjustment command is sent to the control unit of each UPS module through a secure encrypted channel until the control adjustment is completed, and each UPS module sends a confirmation message to the central controller.

[0154] Specifically, first, the power sensor equipped in each UPS module collects the input voltage, input current, output voltage, output current and power factor of the UPS module in real time. These parameters are transmitted to the central controller via industrial Ethernet for subsequent calculation and analysis. For example, the input voltage of a UPS module is 220V, the input current is 100A, the output voltage is 220V, the output current is 95A, and the power factor is 0.9.

[0155] Next, calculate the instantaneous load rate of each UPS module based on its real-time output power and rated capacity. The instantaneous load rate is equal to the real-time output power divided by the rated capacity. For example, if the real-time output power of a UPS module is 20kW and the rated capacity is 25kVA, its instantaneous load rate is 80%.

[0156] Then, based on the instantaneous load rate, the moving average load rate of each UPS module within a preset time period (such as 5 minutes) is calculated. The moving average load rate can be obtained by accumulating the instantaneous load rate at each sampling moment and dividing it by the number of sampling times. For example, if the instantaneous load rates of a UPS module at 10 sampling moments within 5 minutes are 80%, 82%, 81%, 79%, 83%, 80%, 78%, 81%, 82%, and 80%, respectively, then its moving average load rate is 80.6%.

[0157] Determine the average of the moving average load factors of all UPS modules. For example, if there are 3 UPS modules and their moving average load factors are 80.6%, 78.9%, and 82.3%, the average value is 80.6%.

[0158] Calculate the sum of the squares of the difference between the moving average load factor of each UPS module and the average value. Continuing with the above example, the sum of the squares of the difference between the three UPS modules is (80.6%-80.6%)^2+(78.9%-80.6%)^2+(82.3%-80.6%)^2=0+2.89+2.89=5.78.

[0159] Determine the arithmetic mean of the above sum of squares, that is, divide the sum of squares by the number of UPS modules, and the arithmetic mean is 5.78 / 3=1.93.

[0160] Calculate the corresponding square root to get the load balancing index, which is:

[0161] The calculated load balance index is compared with a preset threshold (such as 1.5). When the load balance index is greater than the preset threshold, the power redistribution process is triggered. In this example, 1.39<1.5, so no redistribution is triggered.

[0162] If redistribution is triggered, the power redistribution result needs to be determined. The minimum variance method can be used to make the load rate of each UPS module as close to the average as possible. For example, part of the load of a UPS module with a higher load rate can be transferred to a UPS module with a lower load rate until the load rates of each module are basically the same.

[0163] Based on the power redistribution result, a power adjustment command including a target power value and an adjustment rate is generated. For example, for a UPS module with a load factor of 82.3%, a command to reduce its load factor to 80.6% can be generated, with an adjustment rate of 0.5% per minute.

[0164] The power adjustment command is sent to the control unit of each UPS module through a secure encrypted channel (such as TLS protocol). After receiving the command, the control unit gradually adjusts the output power according to the specified adjustment rate until the target power value is reached.

[0165] After the control adjustment is completed, each UPS module sends a confirmation message to the central controller, indicating that the power adjustment command has been executed. After the central controller receives the confirmation message from all UPS modules, the load balancing control process ends.

[0166] Through the above steps, active current balancing and load sharing control of each module in the UPS system can be achieved, improving the overall efficiency and reliability of the system. This method can monitor the load status of each module in real time, and make timely adjustments when the load is unbalanced to ensure the stable operation of the system.

[0167] In this embodiment, the voltage, current and power factor of the UPS module are collected in real time by power sensors to ensure comprehensive monitoring of the system status and improve data accuracy; the instantaneous load rate and its moving average are calculated to help accurately evaluate the load condition of each UPS module and ensure dynamic adjustment of load balancing; by comparing the load balancing index with a preset threshold, the power redistribution process is automatically triggered to improve the system's adaptability; a power adjustment command including a target power value and an adjustment rate is generated to ensure rapid adjustment when the load is uneven, thereby improving the operating efficiency and stability of the UPS system; the power adjustment command is sent through a secure encrypted channel to ensure the security of data transmission and prevent potential security risks.

[0168] In an optional implementation manner, the power reallocation process includes:

[0169] The UPS modules are sorted from high to low according to the moving average load rate, and starting from the UPS module with the highest load rate, the amount of power transferred to the UPS module with the lowest load rate is calculated, wherein the power amount is calculated as follows:

[0170] Calculate the power difference between the UPS module with the highest load rate and the UPS module with the lowest load rate; calculate half of the rated capacity of the UPS module with the lowest load rate to determine the half capacity; calculate the product of the power difference and the half capacity; compare the power difference with the product, select the minimum value, and determine the power amount;

[0171] The calculation is repeated and the load rate of each UPS module is updated until the load balance index is less than or equal to the preset threshold, and the power redistribution result is determined.

[0172] Specifically, first, the UPS modules are sorted according to their moving average load rates, from high to low. The moving average load rate can be obtained by calculating the average load rate within a certain time window (such as 10 minutes), which can smooth short-term fluctuations and reflect a more stable load state.

[0173] Then, starting from the UPS module with the highest load factor, calculate the amount of power transferred to the UPS module with the lowest load factor. The calculation process is as follows:

[0174] Calculate the power difference between the UPS module with the highest load rate and the UPS module with the lowest load rate. For example, the UPS module with the highest load rate has a load rate of 80%, a rated power of 100kW, and an actual load of 80kW; the UPS module with the lowest load rate has a load rate of 40%, a rated power of 50kW, and an actual load of 20kW. The power difference is 80kW-20kW=60kW.

[0175] Calculate half of the rated capacity of the UPS module with the lowest load factor, i.e. the half capacity value. In this example, the half capacity value is 50kW / 2=25kW.

[0176] Calculate the product of the power difference and the half capacity. In this example, the product is 60kW*25kW / 50kW=30kW.

[0177] The power difference is compared with the product, and the minimum value is selected as the transferable power amount. In this example, the power difference of 60kW is greater than the product of 30kW, so 30kW is selected as the transferable power amount.

[0178] Next, based on the calculated transferable power, the load rates of the two UPS modules are updated. The load of the UPS module with the highest load rate is reduced by 30kW to 50kW, and the load rate is reduced to 50%; the load of the UPS module with the lowest load rate is increased by 30kW to 50kW, and the load rate is increased to 100%.

[0179] Repeat the above calculation process, continue to find the modules with the highest and lowest load rates among the remaining UPS modules, calculate and execute power transfer, until the load balance index of all UPS modules is less than or equal to the preset threshold. The load balance index can be represented by calculating the standard deviation of the load rates of all UPS modules. The smaller the standard deviation, the more balanced the load.

[0180] Assuming that the preset threshold is 5%, after multiple rounds of calculations, the final load rates of each UPS module are 70%, 65%, 68%, and 72%, respectively. At this time, the standard deviation of the load rate is 2.9%, which is less than the preset threshold of 5%, so the power reallocation process is stopped.

[0181] Finally, according to the result of power redistribution, the control system sends corresponding power adjustment instructions to each UPS module to achieve actual load balancing. The control system can achieve power adjustment by adjusting the output voltage or current of the UPS module. The specific adjustment method depends on the control strategy of the UPS.

[0182] Through the above power redistribution process, dynamic load balancing between UPS module groups can be achieved, improving the overall operating efficiency and reliability of the system. This method takes into account the rated capacity differences of UPS modules, avoids the risk of overloading certain modules due to excessive power transfer, and ensures the stability of the system through gradual adjustment.

[0183] In this embodiment, the voltage, current and power factor of the UPS module are collected in real time by power sensors to ensure comprehensive monitoring of the system status and improve data accuracy; the instantaneous load rate and its moving average are calculated to help accurately evaluate the load condition of each UPS module and ensure dynamic adjustment of load balancing; by comparing the load balancing index with a preset threshold, the power redistribution process is automatically triggered to improve the system's adaptability; a power adjustment command including a target power value and an adjustment rate is generated to ensure rapid adjustment when the load is uneven, thereby improving the operating efficiency and stability of the UPS system; the power adjustment command is sent through a secure encrypted channel to ensure the security of data transmission and prevent potential security risks.

[0184] Figure 2 FIG. 1 is a schematic diagram of a UPS capacity setting system for a micro-module computer room based on load prediction according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0185] The first unit is used to obtain the historical load data of the micro-module computer room and build a load prediction model. The load prediction model includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model. The adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, the multi-granularity load changes in the future time period are predicted, and an error compensation mechanism is introduced to adaptively correct the current prediction results to generate a load prediction result containing load demand data at different time scales and different spatial granularities;

[0186] The second unit is used to build a multi-objective UPS capacity configuration model based on the load prediction result, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the optimization multi-objective, integrating dynamic game mechanism and Monte Carlo simulation risk assessment engine, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a dynamically optimized UPS optimal capacity configuration plan;

[0187] The third unit is used to apply the UPS capacity configuration scheme to the UPS system configuration of the micro-module computer room, build a distributed UPS group that meets the needs of different load partitions, and configure load sharing control based on active current sharing. According to the load prediction results and combined with the real-time operating status, the distributed UPS group operating mode and capacity allocation are dynamically adjusted.

[0188] According to a third aspect of the embodiments of the present invention,

[0189] An electronic device is provided, comprising:

[0190] processor;

[0191] a memory for storing processor-executable instructions;

[0192] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0193] According to a fourth aspect of the embodiments of the present invention,

[0194] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0195] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for setting UPS capacity in a micro-module computer room based on load prediction, characterized in that: include: Obtain historical load data of the micro-module computer room and construct a load prediction model, which includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model. An adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, multi-granularity load changes in future time periods are predicted. An error compensation mechanism is introduced to adaptively correct the current prediction results, and a load prediction result containing load demand data at different time scales and different spatial granularities is generated. Based on the load prediction results, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the multi-objective optimization, integrating dynamic game mechanism and Monte Carlo simulation risk assessment engine, constructing a multi-objective UPS capacity configuration model, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a dynamically optimized UPS optimal capacity configuration plan; Apply the UPS optimal capacity configuration scheme to the UPS system configuration of the micro-module computer room, build a distributed UPS group that meets the requirements of different load partitions, and configure load sharing control based on active current sharing. According to the load prediction results and combined with the real-time operation status, dynamically adjust the operation mode and capacity allocation of the distributed UPS group; The historical load data of the micro-module computer room is obtained, and a load prediction model is constructed. The load prediction model includes a regression analysis sub-model, a time series analysis sub-model, and a deep neural network sub-model. An adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, multi-granularity load changes in future time periods are predicted, and an error compensation mechanism is introduced to adaptively correct the current prediction results. The load prediction results generated contain load demand data at different time scales and different spatial granularities, including: Based on the historical load data, a first-layer load prediction model is constructed, and the first-layer load prediction model includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model, wherein the regression analysis sub-model adopts a multivariate linear regression algorithm, and determines the linear relationship between the load and the various influencing factors by modeling the various influencing factors that affect the load; the time series analysis sub-model is constructed based on the ARIMA model, and captures the time dynamic characteristics of the load by analyzing the autocorrelation and stationarity of the load time series; the deep neural network sub-model is constructed based on the LSTM network, captures the dependency of long sequence data, and learns the deep-level characteristics and laws of load changes; the outputs of the regression analysis sub-model, the time series analysis sub-model and the deep neural network sub-model are used as the outputs of the first-layer load prediction model to obtain the first-layer load prediction result; The first-layer load prediction results are used as input features to construct a second-layer load prediction model including a meta-learner. The meta-learner is constructed using an extreme gradient boosting algorithm, and the complex patterns of the first-layer load prediction results are learned by integrating multiple decision trees. In the training process of the meta-learner, an adaptive weight adjustment mechanism based on an online learning method is integrated, and the adaptive weight adjustment mechanism is used to dynamically update the parameters of the meta-learner according to the real-time prediction error of the meta-learner. The updated meta-learner is used to predict the first-layer load prediction results input, obtain the second-layer load prediction results, and generate the load prediction results at the micro-module computer room level. Taking the load prediction result at the micro-module computer room level as input, and using the deep neural network sub-model to decompose the load prediction result at the micro-module computer room level into load prediction results at each rack level; Taking the rack-level load prediction result as input, and using the regression analysis sub-model to further decompose the rack-level load prediction result into a device-level load prediction result; Generate a load forecast result including load demand data at different time scales and different spatial granularities according to the load forecast result at the micromodule room level, the load forecast result at the rack level, and the load forecast result at the device level; Based on the load forecast results, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the multi-objective optimization, the dynamic game mechanism and Monte Carlo simulation risk assessment engine are integrated to build a multi-objective UPS capacity configuration model, adaptively adjust the UPS response time and capacity scheduling strategy, and obtain the dynamically optimized UPS optimal capacity configuration solution including: The weighting factor of the time dimension is introduced, and the first objective function is determined with the goal of minimizing the average utilization of all UPS at all times; the maximum value of the difference between the actual redundancy and the redundancy requirement threshold is calculated, the maximum difference is determined, and the larger value between the maximum difference and 0 is minimized to determine the second objective function; the third objective function is determined with the goal of minimizing the product of the total capacity of all UPS and the unit capacity energy consumption cost coefficient; A dynamic game mechanism is introduced. Based on the first objective function, the second objective function and the third objective function, the corresponding three agents are determined, and the utility function corresponding to each agent is determined. The weight coefficient of the objective function corresponding to each agent is used as the strategy space of each agent. Through the strategy interaction and dynamic adjustment between agents, the Logit dynamic strategy update rule is adopted. Each agent dynamically adjusts its corresponding strategy according to the strategy selection of other agents to maximize its own utility function. Through multiple rounds of iterative games, it converges to the Nash equilibrium point, determines the weight coefficient of the objective function corresponding to each agent, and obtains the initial UPS capacity configuration plan. Performing Monte Carlo simulation risk assessment on the initial UPS capacity configuration scheme, performing multiple random sampling on the load forecast results, generating multiple load scenarios, and for each load scenario, calculating the capacity missing amount and missing duration under the initial UPS capacity configuration scheme, obtaining the corresponding empirical distribution, assessing the risk level of the initial UPS capacity configuration scheme, and introducing tolerance thresholds for the capacity missing amount and missing duration and the corresponding risk threshold as risk constraints; Based on the multi-objective UPS capacity configuration model, dynamic game mechanism and Monte Carlo simulation risk assessment, the UPS capacity dynamic configuration algorithm is determined, the UPS response time and capacity scheduling strategy are adaptively adjusted, and the dynamically optimized UPS optimal capacity configuration plan is obtained.

2. The method according to claim 1, characterized in that The first-layer load prediction results are used as input features to construct a second-layer load prediction model including a meta-learner. The meta-learner is constructed using an extreme gradient boosting algorithm. The complex pattern of the first-layer load prediction results is learned by integrating multiple decision trees, including: The first layer of load prediction results are used as input features, and the data set is divided into a training set and a validation set by cross-validation; Based on the training set, an adaptive weighted sampling method is used to generate a plurality of training subsets with different distributions, and a decision tree is generated based on each of the training subsets. By dynamically adjusting the sampling weight of each of the training subsets, a plurality of decision trees based on different data distributions are obtained to determine a base learner corresponding to the meta learner; Based on each decision tree, a multi-level feature interaction mechanism is constructed to calculate the cross terms and nonlinear transformations of the input features, capture the composite nonlinear relationship in the first-layer load prediction results, and obtain an enhanced expression base learner; According to the residual of the fitting loss function in the negative gradient direction, the structure and parameters of the decision tree corresponding to each enhanced expression base learner are iteratively optimized, and the performance of the enhanced expression base learner is evaluated on the validation set, and the growth of each decision tree is controlled by combining the regularization term and the early stopping mechanism to obtain a generalized base learner; Adopting an adaptive learning rate strategy, the learning rate of each generalized base learner is dynamically adjusted according to the training progress and performance of the decision tree corresponding to each generalized base learner, and the optimal base learner of each decision tree is determined; The optimal base learners are weighted and integrated to determine a meta learner, and the meta learner generates a prediction result of the second-layer load prediction model by learning the first-layer load prediction result.

3. The method according to claim 1, characterized in that The UPS capacity dynamic configuration algorithm includes: Initialize the UPS capacity configuration plan and use the dynamic game mechanism to update the weight coefficients of each optimization target; solve the multi-objective UPS capacity configuration model based on the weight coefficients to obtain a new UPS capacity configuration plan; Conduct Monte Carlo simulation risk assessment on new UPS capacity configuration solutions, calculate the capacity loss probability based on the preset capacity loss tolerance threshold, and calculate the loss duration probability based on the preset loss duration tolerance threshold; Determine whether the probability of capacity loss and the probability of loss duration are less than the risk threshold. If the probability of capacity loss is less than the first risk threshold, and the probability of loss duration is less than the second risk threshold, output the optimal UPS capacity configuration plan; if the probability of capacity loss is greater than or equal to the first risk threshold, or the probability of loss duration is greater than or equal to the second risk threshold, adjust the UPS response time and capacity scheduling strategy, and return to the step of solving the multi-objective UPS capacity configuration model based on the weight coefficient; Repeat the iteration until the preset number of iterations is reached to determine the optimal UPS capacity configuration plan.

4. The method according to claim 1, characterized in that: Active current balancing based load sharing control includes: The power sensor equipped in each UPS module collects the input voltage, input current, output voltage, output current and power factor of the UPS module in real time, and transmits the data to the central controller via industrial Ethernet. Calculate the instantaneous load rate according to the real-time output power and rated capacity of each UPS module, calculate the moving average load rate of each UPS module within a preset time period based on the instantaneous load rate, determine the average value of the moving average load rate, calculate the sum of squares of the difference between the moving average load rate of each UPS module and the average value, determine the arithmetic mean of the sum of squares, calculate the corresponding square root, and determine the load balancing index; Comparing the load balance index with a preset threshold, and when the load balance index is greater than the preset threshold, triggering a power reallocation process and determining a power reallocation result; Based on the power redistribution result, a power adjustment command including the target power value and the adjustment rate is generated, and the power adjustment command is sent to the control unit of each UPS module through a secure encrypted channel until the control adjustment is completed, and each UPS module sends a confirmation message to the central controller.

5. The method according to claim 4, characterized in that The power redistribution process includes: The UPS modules are sorted from high to low according to the moving average load rate, and starting from the UPS module with the highest load rate, the amount of power transferred to the UPS module with the lowest load rate is calculated, wherein the power amount is calculated as follows: Calculate the power difference between the UPS module with the highest load rate and the UPS module with the lowest load rate; calculate half of the rated capacity of the UPS module with the lowest load rate to determine the half capacity; calculate the product of the power difference and the half capacity; compare the power difference with the product, select the minimum value, and determine the power amount; The calculation is repeated and the load rate of each UPS module is updated until the load balance index is less than or equal to the preset threshold, and the power redistribution result is determined.

6. A UPS capacity setting system for a micro-module computer room based on load prediction, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to obtain the historical load data of the micro-module computer room and build a load prediction model. The load prediction model includes a regression analysis sub-model, a time series analysis sub-model and a deep neural network sub-model. The adaptive weight adjustment mechanism is used to dynamically adjust the weight coefficient of each sub-model. By performing multi-dimensional comprehensive modeling on the equipment in the computer room, the multi-granularity load changes in the future time period are predicted, and an error compensation mechanism is introduced to adaptively correct the current prediction results to generate a load prediction result containing load demand data at different time scales and different spatial granularities; The second unit is used to build a multi-objective UPS capacity configuration model based on the load prediction result, taking UPS capacity utilization, redundancy requirements and energy consumption cost combination as the optimization multi-objective, integrating dynamic game mechanism and Monte Carlo simulation risk assessment engine, adaptively adjusting UPS response time and capacity scheduling strategy, and obtaining a dynamically optimized UPS optimal capacity configuration plan; The third unit is used to apply the UPS capacity configuration scheme to the UPS system configuration of the micro-module computer room, build a distributed UPS group that meets the needs of different load partitions, and configure load sharing control based on active current sharing. According to the load prediction results and combined with the real-time operating status, the distributed UPS group operating mode and capacity allocation are dynamically adjusted.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images