Energy management method and system of a machine room optical storage and charging integrated system

Abstract

The application provides an energy management method and system of a machine room optical storage and charging integrated system, comprising obtaining machine room state information and external environment information; calculating a running flexibility index of the machine room based on a total power margin of adjustable non-critical IT load and an adjustment duration, and selecting a quantile point in a probabilistic prediction distribution of the photovoltaic power as a deterministic photovoltaic output prediction value according to the flexibility index; constructing a running cost model comprising a power purchase cost in a scheduling period, a storage unit life loss cost and a power adjustment cost of the non-critical IT load; the life loss cost is represented according to a state of charge and a health state of the storage unit, and a penalty weight in the power adjustment cost is adjusted according to the health state of the storage unit; and under the constraint of meeting a power demand of the critical IT load, a charging and discharging power of the storage unit and a power adjustment scheme of the non-critical IT load are obtained by joint optimization.

Description

Technical Field

[0001] This application belongs to the field of management, and in particular relates to an energy management method and system for an integrated optical storage and charging system for a data center. Background Technology

[0002] Data center energy management methods aim for optimal economic efficiency by rationally scheduling the charging and discharging behavior of energy storage units based on the grid's time-of-use pricing. These methods rely on deterministic predictions of future photovoltaic (PV) output, using a prediction curve as the basis for energy scheduling. However, PV output is intermittent and uncertain, and point predictions cannot accurately represent the true fluctuation range. When actual PV output is significantly lower or higher than the prediction, the scheduling strategy based on this prediction will be biased. In constructing optimization models for energy management systems, most methods only consider electricity purchase costs, neglecting the lifespan depreciation costs of energy storage units during frequent charging and discharging. While some methods consider energy storage depreciation, they fail to incorporate real-time state of charge and health factors into the cost representation, potentially leading to overuse and accelerated performance degradation. IT loads within data centers have different service levels, with some non-critical IT loads having adjustable power potential within a certain range. However, existing technologies do not jointly optimize energy storage scheduling with the power regulation of non-critical IT loads, nor do they establish a cost link between the two. Furthermore, the forecast of photovoltaic output is inaccurate, leading to inaccurate subsequent adjustments. Existing technologies lack a method to weigh whether to prioritize energy storage or load adjustment based on the health status of the energy storage unit. Summary of the Invention

[0003] This invention proposes an energy management method for an integrated optical-storage-charging system in a data center, addressing the problem that existing technologies cannot prioritize energy storage or load adjustment based on the health status of the energy storage unit. The method includes: The system acquires data center status information and external environment information. The data center status information includes the service level of the IT load, the state of charge and health status of the energy storage units, and the external environment information includes historical photovoltaic power data, solar intensity and temperature data from weather forecasts, and time-of-use electricity prices from the power grid. Based on an improved quantile regression neural network and the external environment information, a probabilistic predicted distribution of photovoltaic power is obtained. Based on the total power margin and adjustment time of the adjustable non-critical IT load, the computer room's operational flexibility index is determined, and a quantile is selected in the probabilistic prediction distribution of the photovoltaic power according to the flexibility index as a deterministic photovoltaic output prediction value. Construct an operating cost model that includes the cost of electricity purchase during the scheduling cycle, the cost of energy storage unit lifespan loss, and the power regulation cost of non-critical IT loads; With the goal of minimizing operating costs, and under the constraint of meeting the power requirements of critical IT loads, a joint optimization solution is obtained to obtain the charging and discharging power of the energy storage unit and the power regulation scheme of non-critical IT loads.

[0004] Furthermore, this invention also relates to an energy management system for an integrated optical storage and charging system for a data center, comprising the following modules: The acquisition module is used to acquire data center status information and external environment information. The data center status information includes the service level of the IT load, the state of charge and health status of the energy storage unit, and the external environment information includes historical photovoltaic power data, light intensity and temperature data in the weather forecast, and the time-of-use electricity price of the power grid. Based on the improved quantile regression neural network and the external environment information, a probabilistic predicted distribution of photovoltaic power is obtained. The selection module is used to select a quantile in the probabilistic prediction distribution of photovoltaic power based on the total power margin and adjustment time of the adjustable non-critical IT load, the operational flexibility index of the computer room, and the flexibility index as a deterministic photovoltaic output prediction value. Establish a module to build an operating cost model that includes the cost of electricity purchase during the scheduling cycle, the cost of energy storage unit lifespan loss, and the power regulation cost of non-critical IT loads; The solution module is used to jointly optimize and solve for the charging and discharging power of the energy storage unit and the power regulation scheme of non-critical IT loads, with the goal of minimizing the operating cost and under the constraint of meeting the power requirements of critical IT loads.

[0005] Compared with the prior art, the present invention has the following beneficial effects: 1) In the probabilistic prediction distribution calculation of photovoltaic power, this invention employs a comprehensive loss function with a quantile crossover penalty term, reducing the possibility of quantile crossover problems; furthermore, the discrete prediction results output by the quantile regression neural network are smoothed into a continuous probability density function using KDE, allowing the extraction of power prediction values ​​at any quantile. 2) By representing the operational flexibility of the data center itself and selecting deterministic predicted values ​​from the probabilistic prediction distribution of photovoltaic power, the energy dispatch scheme's ability to cope with the uncertainty of photovoltaic output is improved. A cost model incorporating energy storage lifetime loss and load regulation is established, which integrates the real-time state of charge and health status of energy storage units into the representation of lifetime loss costs. A correlation mechanism between load regulation costs and energy storage lifetime loss costs is established, which tends to regulate non-critical IT loads when energy storage health is low, and prioritizes the use of energy storage when its health is good, thus achieving rational utilization of both resources.

[0006] This invention optimizes the minimization of operating costs, thereby reducing the electricity purchase expenditure of the data center while ensuring the power supply of critical IT loads, and slowing down the performance degradation of energy storage units, thus improving the long-term economic benefits of the system. Attached Figure Description

[0007] Figure 1 A flowchart of the first embodiment; Figure 2 This is a schematic diagram of information acquisition and system architecture. Figure 3 This is a schematic diagram illustrating the selection of photovoltaic prediction quantiles based on flexible indicators. Figure 4 A schematic diagram illustrating the relationship between penalty weights and the health status of energy storage; Figure 5 This is a schematic diagram of the joint optimization solution framework. Detailed Implementation

[0008] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0009] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0010] In the first embodiment, the present invention proposes an energy management method for an integrated optical storage and charging system for a data center, such as... Figure 1 ,include: S1. Obtain data center status information and external environment information. The data center status information includes the service level of the IT load, the state of charge and health status of the energy storage unit, and the external environment information includes historical photovoltaic power data, light intensity and temperature data in the weather forecast, and the time-of-use electricity price of the power grid. Based on the improved quantile regression neural network and the external environment information, obtain the probabilistic prediction distribution of photovoltaic power. The system collects power data from each IT load in real time through the data center monitoring system and reads the business level of each load from a pre-set business level library. For example, online transaction processing loads with high real-time requirements are defined as critical loads, while data backup and model training loads that can be executed with delays are defined as non-critical loads. The system obtains the current State of Charge (SOC) and State of Health (SOH) values ​​of the energy storage units through the Battery Management System (BMS). Solar irradiance and temperature weather forecast data for the next 24 hours are obtained by calling a third-party meteorological service interface. This data is input into a pre-trained quantile regression neural network model, which outputs photovoltaic power prediction values ​​corresponding to multiple quantiles from 0 to 1 at each future scheduling time, collectively forming a probabilistic prediction distribution of photovoltaic power. The system obtains the time-of-use electricity price table for the next 24 hours, including specific prices for peak, flat, and valley periods, by accessing the official website or dedicated data interface of the power grid company. Figure 2 .

[0011] In an optional embodiment, the probabilistic prediction distribution of photovoltaic power obtained based on the improved quantile regression neural network and the external environmental information specifically includes: The external environment information is normalized, and the Pearson correlation coefficient between each external environment feature and historical photovoltaic power is calculated. Feature vectors with correlation coefficients greater than a preset threshold are selected as network inputs. A quantile regression neural network model based on the combination of temporal convolutional network and multi-head self-attention mechanism is constructed. The quantile regression neural network model uses a temporal convolutional network as the feature extraction layer and introduces a multi-head self-attention mechanism layer after the TCN layer. The model is trained using a comprehensive loss function with a quantile cross-penalty term; the comprehensive loss function consists of a bouncing loss function and a cross-penalty term. The filtered feature vectors are input into the trained model, which outputs photovoltaic power prediction values ​​at multiple discrete quantiles. The prediction values ​​at the discrete quantiles are then fitted using a nonparametric kernel density estimation method to generate a continuous photovoltaic power probability density function, which serves as the probabilistic prediction distribution of photovoltaic power.

[0012] The acquired external environmental information is normalized, and the Pearson correlation coefficient between each external environmental feature and historical photovoltaic power is calculated. Feature vectors with correlation coefficients greater than a preset threshold are selected as model inputs to reduce the impact of redundant features on prediction accuracy. The constructed quantile regression neural network model uses a temporal convolutional network (TCN) as the feature extraction layer to obtain local temporal dependencies in the photovoltaic power time series. A multi-head self-attention mechanism layer is introduced at the TCN output to enhance the model's adaptive weighting capability for features at different time scales. The model output layer outputs the photovoltaic power prediction value at each preset quantile level τ (τ∈(0,1)), thereby obtaining a set of discrete quantile prediction results at the same time. During model training, historical photovoltaic power data is used as a supervision signal, and the network parameters are iteratively updated through the backpropagation algorithm to gradually approximate the true power distribution at each quantile level. More specifically, during the model training phase, a comprehensive loss function with a quantile cross-penalty term is used as the optimization objective, where the bouncing loss function is used to measure the accuracy of the prediction value at each quantile level. For a given quantile level τ, the corresponding bouncing loss function is:

[0013] Where y represents the actual photovoltaic power value. This represents the predicted value at that quantile. To avoid overlap or order violations between predictions from different quantiles, a quantile crossover penalty term is introduced into the bouncing loss function. When the predicted value at a lower quantile is higher than that at a higher quantile, an additional penalty is applied. The comprehensive loss function is: . For the set of quantiles, The penalty coefficient is used. After model training, the selected feature vectors are input into the quantile regression neural network model to obtain photovoltaic power prediction values ​​corresponding to multiple discrete quantiles. To construct a continuous probabilistic prediction distribution from the discrete quantile prediction results, a nonparametric kernel density estimation method is used to fit the above quantile prediction values. The nonparametric kernel density estimation is a probability density estimation method that does not rely on a preset distribution form. It obtains a continuous probability density function of the target random variable by superimposing a kernel function at each sample point and performing smoothing. By performing kernel density estimation on the predicted values ​​of multiple quantiles, a continuous probability density function of photovoltaic power at the prediction time can be generated.

[0014] In an alternative embodiment, the probabilistic prediction distribution of photovoltaic power is determined as follows: A quantile regression neural network model is used as input, along with historical photovoltaic power data, solar irradiance and temperature data from weather forecasts. The model outputs a probabilistic predicted distribution of photovoltaic power at each preset time step within a preset time period, between the first preset quantile and the second preset quantile.

[0015] A quantile regression neural network model is constructed and trained. The structure includes a recurrent neural network layer, such as a long short-term memory network, for detecting time-series features, and one or more fully connected layers. The model uses quantile loss, also known as the bouncing ball loss function, to predict different quantiles. For example, the model's output layer can be set to 9 neurons, corresponding to quantiles from 0.1, 0.2, up to 0.9, to output the predicted value of photovoltaic power at the stated probability level.

[0016] During forecasting, measured photovoltaic (PV) power values ​​are collected every 15 minutes over the past 24 hours, along with weather forecasts for the next 24 hours, including solar irradiance and temperature for each 15-minute step. This sequence data is then input into a trained model. For example, to predict PV power for the next 15 minutes, the model receives historical power data sequences such as 120 kW, 125 kW, and future weather data such as solar irradiance of 800 W / m² and temperature of 25°C. It outputs a vector, such as [80, 88, 95, 102, 110, 115, 120, 124, 128] kW, which represents the predicted PV power values ​​from the 0.1 to 0.9 quantiles for that time step, forming a probabilistic distribution of PV power prediction. This process is repeated for each 15-minute step over the next 24 hours to obtain the complete probabilistic prediction result.

[0017] The model structure's input layer receives concatenated multidimensional time-series data. A Long Short-Term Memory (LSTM) network layer with, for example, 128 units is used to extract time dependencies. Feature enhancement is achieved through two fully connected layers, each containing 64 neurons and using linear rectified unit (RC) activation functions. The output layer has 9 neurons and does not use an activation function, corresponding to quantile predictions from 0.1 to 0.9. The training set is constructed using at least two years of historical data, including actual photovoltaic (PV) power every 15 minutes, and corresponding meteorological data such as light intensity and temperature. This creates a large number of input-output sample pairs. Each sample's input is historical data from the past 96 time steps (24 hours), and its output is the actual PV power for the next time step. The training set is divided into mini-batch datasets, and a moment estimation optimizer is used. The quantile loss function is minimized using the backpropagation algorithm. Iteratively update network weights, where , It is the actual value. It is the predicted value of the quantile q. It is an indicator function. The training process will continue for multiple rounds until the model's loss on the validation set converges.

[0018] S2, based on the total power margin and adjustment time of the adjustable non-critical IT load, the computer room operation flexibility index, and according to the flexibility index, a quantile in the probabilistic prediction distribution of the photovoltaic power is selected as the deterministic photovoltaic output prediction value. The total power regulation margin at the current moment is obtained by summing the differences between the current power and the minimum allowable operating power of all non-critical IT loads. Non-critical IT loads include, but are not limited to, time-delayed computing tasks, batch processing tasks, and non-real-time data processing tasks. Critical IT loads include, but are not limited to, real-time business processing tasks, core business systems, and computing tasks with high service continuity requirements. Combined with the maximum allowable interruption or delay duration, the total regulated power available within the scheduling cycle is calculated as a core indicator of the data center's operational flexibility. A mapping relationship is established between the flexibility indicator and the photovoltaic (PV) prediction percentile, such as a linear or piecewise linear function. When the data center's flexibility indicator is high, it indicates a strong ability to withstand fluctuations in PV output. In this case, a higher percentile, such as the 75th percentile, is selected as the predicted PV output value to pursue higher PV absorption and returns. When the data center's flexibility indicator is low, it indicates limited regulation capacity. In this case, a lower conservative percentile, such as the 25th percentile, is selected as the predicted PV output value to prioritize power supply reliability.

[0019] In an optional embodiment, the operational flexibility index of the computer room, based on the total power margin and adjustment duration of the adjustable non-critical IT load, includes: Through formula The calculation is performed, where F is the flexibility index. The total power reduction margin for all current adjustable non-critical IT loads is adjusted downwards. The maximum allowable continuous adjustment duration, This refers to the rated capacity of the energy storage unit.

[0020] Identify all non-critical IT equipment in the data center that can be interrupted or run at reduced frequencies, such as server clusters used for offline rendering or big data analytics. Assuming there are 20 servers, each with an average operating power of 2 kilowatts, what is the total power reduction margin if business needs can be completely suspended? It is 40 kilowatts.

[0021] Maximum continuous adjustment duration Determined by the service level agreement or operational strategy, this refers to the maximum duration for which non-critical tasks can be continuously suspended without materially impacting business operations. For example, if batch processing tasks must be completed before 4:00 AM daily, and there is currently a 3-hour buffer period, then the maximum allowed continuous adjustment time is... It can be set to 2 hours. Obtain the rated capacity of the energy storage unit. This is a fixed parameter, for example, the rated capacity of the battery energy storage system equipped in a data center is 200 kWh. Substituting this value into the formula, the flexibility index F = 0.4 is calculated. The dimensionless value 0.4 is the current operational flexibility index of the data center.

[0022] To generate a deterministic prediction curve for optimized scheduling, in an optional embodiment, selecting a quantile in the probabilistic prediction distribution of photovoltaic power based on the flexibility index includes: When the flexibility index F is greater than or equal to the first flexibility threshold, the first target quantile is selected; When the flexibility index F is between the second flexibility threshold and the first flexibility threshold, the second target quantile is selected. When the flexibility index F is less than the second flexibility threshold, the third target quantile is selected as the deterministic photovoltaic output prediction value.

[0023] Specifically, a set of flexible thresholds and corresponding target quantiles are preset. For example, the first flexible threshold is set to 0.8, and the second flexible threshold is set to 0.3; simultaneously, the first target quantile is set to 0.7 to represent an optimistic prediction, the second target quantile to 0.5 to represent a neutral prediction, and the third target quantile to 0.2 to represent a conservative prediction. These values ​​are set based on the operator's risk tolerance and operational experience.

[0024] During execution, assume that F is calculated to be 0.4 using the aforementioned method. In this example, 0.3 < 0.4 and 0.4 < 0.8, therefore F falls between the second and first flexible thresholds. According to the rules, the second target quantile, i.e., the 0.5 quantile, should be selected at this point. From the previously obtained probabilistic photovoltaic power prediction distribution, the 0.5 quantile prediction values ​​for all future time steps are extracted to form a single time series curve. For example, if the 0.5 quantile prediction value for the next three hours is [110, 112, 108] kW, then this curve is used as the deterministic photovoltaic output prediction value to construct the power balance constraint of the equipment room within the scheduling cycle, in order to determine the purchased power and the charging and discharging power of the energy storage unit. If F is calculated to be 0.2, then the conservative 0.2 quantile curve will be selected, such as... Figure 3 .

[0025] S3, construct an operating cost model that includes the cost of electricity purchase during the scheduling cycle, the cost of energy storage unit lifespan loss, and the power regulation cost of non-critical IT loads; The cost of electricity purchase is the sum of the products of the electricity purchased from the grid in each time period and the corresponding electricity price for that time period. The lifespan loss cost of energy storage units is represented by taking into account real-time conditions. For example, a cycle lifespan loss model based on rainflow counting can be established, where the loss cost per cycle is related to the depth of discharge, the current state of equilibrium (SOH) value, and the state of charge (SOC) range. The lower the SOH value, the higher the loss cost per cycle; when the SOC is in extremely high or low ranges, such as below 20% or above 90%, charging and discharging behavior will generate additional loss cost factors. The power regulation cost of non-critical IT loads is the product of the regulation power and the penalty weight, which is inversely proportional to the lifespan loss cost of the energy storage unit. For example, when the calculated energy storage lifespan loss cost is high, such as when the SOH is below the 85% threshold, the penalty weight for load regulation is set to a lower value, encouraging priority to load regulation rather than using energy storage. When the energy storage is in good health, the penalty weight for load regulation is set to a higher value, prioritizing the use of energy storage for peak shaving and valley filling. Figure 4 .

[0026] To assess the economic losses caused by a specific charge-discharge strategy to the lifespan of an energy storage unit, in one optional embodiment, the lifespan loss cost is calculated as follows: The rainflow counting method is used to count the number of charge-discharge cycles and the depth of discharge of the energy storage unit within a scheduling cycle. Combined with the current state of health (SOH) value of the energy storage unit, the corresponding life loss is calculated by looking up the preset battery cycle life curve, and then multiplied by the replacement cost per unit capacity to obtain the life loss cost.

[0027] Obtain the State of Charge (SOC) change sequence within a scheduling cycle. For example, a 24-hour scheduling scheme to be evaluated might produce a simplified SOC sequence percentage as follows [80, 50, 70, 40, 80]. Analyze this SOC sequence using rainflow counting, an algorithm that can identify complete and partial cycles and their depths within the sequence. For the above sequence, rainflow counting identifies a cycle from 70 to 50 and back to 70 with a Depth of Discharge (DOD) of 20 percentage points, and a cycle from 80 to 40 and back to 80 with a DOD of 40 percentage points.

[0028] Calculating lifetime loss requires consulting a pre-defined battery cycle life curve, typically provided by the manufacturer, which represents the total number of battery cycles under different DOD (Device Optimization) values. Assuming the curve indicates that at the current 90% SOH (State of Health), a 20% DOD results in 8000 cycles, and a 40% DOD results in 3000 cycles, then the lifetime loss from a single 20% DOD cycle is 1 divided by 8000 of the total lifetime, and the loss from a single 40% DOD cycle is 1 divided by 3000. The total lifetime loss is the sum of these two fractions. Multiplying the calculated total lifetime loss fraction by the total replacement cost of the energy storage unit, for example, $50,000, yields the lifetime loss cost resulting from this scheduling scheme.

[0029] In an optional embodiment, the penalty weight in the power regulation cost is calculated as follows: Through formula Calculate the penalty weight W in the power regulation cost, where k is a preset regulation coefficient. This represents the health status value of the energy storage unit at the start of the current scheduling cycle.

[0030] Specifically, the healthier the energy storage unit, the higher its potential value, and its extended lifespan should be used cautiously, thus incurring a high penalty weight. When the unit's health deteriorates, the usage penalty is appropriately reduced. The adjustment coefficient k is a hyperparameter determined based on a trade-off between economic benefits and equipment protection strategies. For example, k might be set to 50.

[0031] At the beginning of each scheduling cycle, the health status of the energy storage units is obtained from the battery management system. SOH is a value between 0 and 1, where 1 represents a completely new state. Assuming that at the start of a new scheduling cycle, the measurement is... A value of 0.95 indicates a battery health level of 95%. (The last part, "k and," appears to be a separate, unrelated statement.) Substituting into the formula yields W=47.5. This calculated weight W will be used to construct the optimization objective function, for example, by multiplying it by the absolute value of the power regulation of non-critical IT loads, thereby imposing a corresponding cost penalty on drastic power regulation during optimization. If, after several months, SOH decreases to 0.8, the new weight will be W=40, with a reduced penalty, favoring the use of energy storage.

[0032] S4. With the goal of minimizing the operating cost, and under the constraint of meeting the power requirements of critical IT loads, the charging and discharging power of the energy storage unit and the power regulation scheme of non-critical IT loads are obtained through joint optimization.

[0033] The objective function is the sum of electricity purchase costs, energy storage lifetime loss costs, and non-critical IT load adjustment costs at all times within a scheduling period, such as the next 24 hours. Constraints include: power balance constraints at each time point, i.e., the sum of photovoltaic output, grid-purchased power, and energy storage discharge power equals the sum of critical IT load, non-critical IT load, and energy storage charging power; energy storage unit operation constraints, including maintaining the state of charge (SOC) within a safe range and ensuring that charging and discharging power does not exceed the rated maximum power; grid interaction power constraints, i.e., the power purchased from the grid cannot exceed the transformer capacity limit; and non-critical IT load adjustment constraints, i.e., the power reduction cannot exceed the total power adjustment margin. These objective functions and constraints are constructed into a mixed-integer linear programming problem, solved using commercial solvers such as Gurobi or CPLEX, to obtain the optimal energy storage charging and discharging power values ​​and the optimal power adjustment amounts for non-critical IT loads for each scheduling period within the next 24 hours. Figure 5 .

[0034] In an optional embodiment, the optimization objective of minimizing operating costs, while satisfying the power requirements of critical IT loads, involves jointly optimizing the charging and discharging power of the energy storage unit and the power regulation scheme for non-critical IT loads, including: The optimization problem is constructed as a mixed-integer linear programming model and solved using an optimization solver; Set the scheduling period to a preset duration and the time step to a preset time step. Solve and output the set value of the energy storage unit's charging and discharging power and the percentage reduction of the power of each non-critical IT load for each time step in the next scheduling period.

[0035] Construct the components of a mixed-integer linear programming model. The objective function is to minimize the total operating cost over the next 24 hours, which is the sum of grid power purchase cost, energy storage lifetime depreciation cost, and power regulation cost of non-critical IT loads. Decision variables include grid power purchase, energy storage charging power, energy storage discharging power, and power downscaling coefficients for each 15-minute time step, where the charging and discharging states of energy storage can be represented by binary variables to ensure that they do not occur simultaneously.

[0036] A series of linear constraints are set, including power balance constraints at each time step (the sum of photovoltaic output, grid-purchased power, and energy storage discharge power must equal the sum of critical IT load power, non-critical IT load power, and energy storage charging power); operational constraints for energy storage units, such as upper and lower limits of state of charge (SOC), maximum charge / discharge power limits, and equations for SOC changes with charge / discharge); 100% satisfaction of critical IT load requirements; and constraints that the power reduction ratio for non-critical IT loads cannot exceed the maximum allowable value. The completed model is converted according to the input formats supported by the Gurobi / CPLEX solver, such as Python Gurobi API and OPL language. The solution parameters are set to relative optimal gap ≤ 0.01% and maximum solution time ≤ 300s. The solution is then submitted. The solver outputs the energy storage charge / discharge power setpoints and the percentage reduction in non-critical IT load power at each time step within the next scheduling cycle, serving as the final scheduling instruction.

[0037] In the second embodiment, the present invention also proposes an energy management system for an integrated optical storage and charging system for a data center, comprising the following modules: The acquisition module is used to acquire data center status information and external environment information. The data center status information includes the service level of the IT load, the state of charge and health status of the energy storage unit, and the external environment information includes historical photovoltaic power data, light intensity and temperature data in the weather forecast, and the time-of-use electricity price of the power grid. Based on the improved quantile regression neural network and the external environment information, a probabilistic predicted distribution of photovoltaic power is obtained. The selection module is used to select a quantile in the probabilistic prediction distribution of photovoltaic power based on the total power margin and adjustment time of the adjustable non-critical IT load, the operational flexibility index of the computer room, and the flexibility index as a deterministic photovoltaic output prediction value. Establish a module to build an operating cost model that includes the cost of electricity purchase during the scheduling cycle, the cost of energy storage unit lifespan loss, and the power regulation cost of non-critical IT loads; The solution module is used to jointly optimize and solve for the charging and discharging power of the energy storage unit and the power regulation scheme of non-critical IT loads, with the goal of minimizing the operating cost and under the constraint of meeting the power requirements of critical IT loads.

[0038] In an optional embodiment, the probabilistic prediction distribution of photovoltaic power obtained based on the improved quantile regression neural network and the external environmental information specifically includes: The external environment information is normalized, and the Pearson correlation coefficient between each external environment feature and historical photovoltaic power is calculated. Feature vectors with correlation coefficients greater than a preset threshold are selected as network inputs. A quantile regression neural network model based on the combination of temporal convolutional network and multi-head self-attention mechanism is constructed. The quantile regression neural network model uses a temporal convolutional network as the feature extraction layer and introduces a multi-head self-attention mechanism layer after the TCN layer. The model is trained using a comprehensive loss function with a quantile cross-penalty term; the comprehensive loss function consists of a bouncing loss function and a cross-penalty term. The filtered feature vectors are input into the trained model, which outputs photovoltaic power prediction values ​​at multiple discrete quantiles. The prediction values ​​at the discrete quantiles are then fitted using a nonparametric kernel density estimation method to generate a continuous photovoltaic power probability density function, which serves as the probabilistic prediction distribution of photovoltaic power.

[0039] In an optional embodiment, the operational flexibility index of the computer room based on the total power margin and adjustment duration of the adjustable non-critical IT load includes: Through formula The calculation is performed, where F is the flexibility index. The total power reduction margin for all current adjustable non-critical IT loads is adjusted downwards. The maximum allowable continuous adjustment duration, This refers to the rated capacity of the energy storage unit.

[0040] In an optional embodiment, selecting a quantile in the probabilistic prediction distribution of photovoltaic power based on the flexibility index includes: When the flexibility index F is greater than or equal to the first flexibility threshold, the first target quantile is selected; When the flexibility index F is between the second flexibility threshold and the first flexibility threshold, the second target quantile is selected. When the flexibility index F is less than the second flexibility threshold, the third target quantile is selected as the deterministic photovoltaic output prediction value.

[0041] In an optional embodiment, the lifetime attrition cost is calculated as follows: The rainflow counting method is used to count the number of charge-discharge cycles and the depth of discharge of the energy storage unit within a scheduling cycle. Combined with the current state of health (SOH) value of the energy storage unit, the corresponding life loss is calculated by looking up the preset battery cycle life curve, and then multiplied by the replacement cost per unit capacity to obtain the life loss cost.

[0042] In an optional embodiment, the penalty weight in the power regulation cost is calculated as follows: Through formula Calculate the penalty weight W in the power regulation cost, where k is a preset regulation coefficient. This represents the health status value of the energy storage unit at the start of the current scheduling cycle.

[0043] In an optional embodiment, the optimization objective of minimizing operating costs, while satisfying the power requirements of critical IT loads, involves jointly optimizing the charging and discharging power of the energy storage unit and the power regulation scheme for non-critical IT loads, including: The optimization problem is constructed as a mixed-integer linear programming model and solved using an optimization solver; Set the scheduling period to a preset duration and the time step to a preset time step. Solve and output the set value of the energy storage unit's charging and discharging power and the percentage reduction of the power of each non-critical IT load for each time step in the next scheduling period.

[0044] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0045] The functional modules shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0046] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0047] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0048] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. An energy management method for an integrated optical storage and charging system in a data center, characterized in that, Includes the following steps: The system acquires data center status information and external environment information. The data center status information includes the service level of the IT load, the state of charge and health status of the energy storage units, and the external environment information includes historical photovoltaic power data, solar intensity and temperature data from weather forecasts, and time-of-use electricity prices from the power grid. Based on an improved quantile regression neural network and the external environment information, a probabilistic predicted distribution of photovoltaic power is obtained. Based on the total power margin and adjustment time of the adjustable non-critical IT load, the computer room's operational flexibility index is determined, and a quantile is selected in the probabilistic prediction distribution of the photovoltaic power according to the flexibility index as a deterministic photovoltaic output prediction value. Construct an operating cost model that includes the cost of electricity purchase during the scheduling cycle, the cost of energy storage unit lifespan loss, and the power regulation cost of non-critical IT loads; With the goal of minimizing operating costs, and under the constraint of meeting the power requirements of critical IT loads, a joint optimization solution is obtained to obtain the charging and discharging power of the energy storage unit and the power regulation scheme of non-critical IT loads.

2. The energy management method as described in claim 1, characterized in that, The probabilistic prediction distribution of photovoltaic power obtained based on the improved quantile regression neural network and the external environmental information is as follows: The external environment information is normalized, and the Pearson correlation coefficient between each external environment feature and historical photovoltaic power is calculated. Feature vectors with correlation coefficients greater than a preset threshold are selected as network inputs. A quantile regression neural network model based on the combination of temporal convolutional network and multi-head self-attention mechanism is constructed. The quantile regression neural network model uses a temporal convolutional network as the feature extraction layer and introduces a multi-head self-attention mechanism layer after the TCN layer. The model is trained using a comprehensive loss function with a quantile cross-penalty term; the comprehensive loss function consists of a bouncing loss function and a cross-penalty term. The filtered feature vectors are input into the trained model, which outputs photovoltaic power prediction values ​​at multiple discrete quantiles. The prediction values ​​at the discrete quantiles are then fitted using a nonparametric kernel density estimation method to generate a continuous photovoltaic power probability density function, which serves as the probabilistic prediction distribution of photovoltaic power.

3. The energy management method as described in claim 1, characterized in that, The operational flexibility indicators of the computer room, based on the total power margin and adjustment duration of adjustable non-critical IT loads, include: Through formula The calculation is performed, where F is the flexibility index. The total power reduction margin for all current adjustable non-critical IT loads is adjusted downwards. The maximum allowable continuous adjustment duration, This refers to the rated capacity of the energy storage unit.

4. The energy management method as described in claim 1, characterized in that, The step of selecting a quantile in the probabilistic prediction distribution of photovoltaic power based on the flexibility index includes: When the flexibility index F is greater than or equal to the first flexibility threshold, the first target quantile is selected; When the flexibility index F is between the second flexibility threshold and the first flexibility threshold, the second target quantile is selected. When the flexibility index F is less than the second flexibility threshold, the third target quantile is selected as the deterministic photovoltaic output prediction value.

5. The energy management method as described in claim 1, characterized in that, The calculation method for the lifespan depletion cost is as follows: The rainflow counting method is used to count the number of charge-discharge cycles and the depth of discharge of the energy storage unit within a scheduling cycle. Combined with the current state of health (SOH) value of the energy storage unit, the corresponding life loss is calculated by looking up the preset battery cycle life curve, and then multiplied by the replacement cost per unit capacity to obtain the life loss cost.

6. The energy management method as described in claim 1, characterized in that, The penalty weight in the power regulation cost is calculated as follows: Through formula Calculate the penalty weight W in the power regulation cost, where k is a preset regulation coefficient. This represents the health status value of the energy storage unit at the start of the current scheduling cycle.

7. The energy management method as described in claim 1, characterized in that, The optimization objective is to minimize the operating cost. Under the constraint of meeting the power requirements of critical IT loads, the joint optimization solution obtains the charging and discharging power of the energy storage unit and the power regulation scheme of non-critical IT loads, including: The optimization problem is constructed as a mixed-integer linear programming model and solved using an optimization solver; Set the scheduling period to a preset duration and the time step to a preset time step. Solve and output the set value of the energy storage unit's charging and discharging power and the percentage reduction of the power of each non-critical IT load for each time step in the next scheduling period.

8. An energy management system for an integrated photovoltaic, energy storage, and charging system in a data center, characterized in that, Includes the following modules: The acquisition module is used to acquire data center status information and external environment information. The data center status information includes the service level of the IT load, the state of charge and health status of the energy storage unit, and the external environment information includes historical photovoltaic power data, light intensity and temperature data in the weather forecast, and the time-of-use electricity price of the power grid. Based on the improved quantile regression neural network and the external environment information, a probabilistic predicted distribution of photovoltaic power is obtained. The selection module is used to select a quantile in the probabilistic prediction distribution of photovoltaic power based on the total power margin and adjustment time of the adjustable non-critical IT load, the operational flexibility index of the computer room, and the flexibility index as a deterministic photovoltaic output prediction value. Establish a module to build an operating cost model that includes the cost of electricity purchase during the scheduling cycle, the cost of energy storage unit lifespan loss, and the power regulation cost of non-critical IT loads; The solution module is used to jointly optimize and solve for the charging and discharging power of the energy storage unit and the power regulation scheme of non-critical IT loads, with the goal of minimizing the operating cost and under the constraint of meeting the power requirements of critical IT loads.

9. The energy management system as described in claim 8, characterized in that, The probabilistic prediction distribution of photovoltaic power obtained based on the improved quantile regression neural network and the external environmental information is as follows: The external environment information is normalized, and the Pearson correlation coefficient between each external environment feature and historical photovoltaic power is calculated. Feature vectors with correlation coefficients greater than a preset threshold are selected as network inputs. A quantile regression neural network model based on the combination of temporal convolutional network and multi-head self-attention mechanism is constructed. The quantile regression neural network model uses a temporal convolutional network as the feature extraction layer and introduces a multi-head self-attention mechanism layer after the TCN layer. The model is trained using a comprehensive loss function with a quantile cross-penalty term; the comprehensive loss function consists of a bouncing loss function and a cross-penalty term. The filtered feature vectors are input into the trained model, which outputs photovoltaic power prediction values ​​at multiple discrete quantiles. The prediction values ​​at the discrete quantiles are then fitted using a nonparametric kernel density estimation method to generate a continuous photovoltaic power probability density function, which serves as the probabilistic prediction distribution of photovoltaic power.

10. The energy management system as described in claim 8, characterized in that, The operational flexibility indicators of the computer room, based on the total power margin and adjustment duration of adjustable non-critical IT loads, include: Through formula The calculation is performed, where F is the flexibility index. The total power reduction margin for all current adjustable non-critical IT loads is adjusted downwards. The maximum allowable continuous adjustment duration, This refers to the rated capacity of the energy storage unit.

Images