CN116739292A - Energy optimization scheduling method, system and storage medium of data center - Google Patents

Abstract

The invention provides an energy optimization scheduling method, system and storage medium for a data center. The method includes: outputting the photovoltaic power generation output power predicted in the past day based on numerical weather prediction data and pre-trained BiLSTM prediction network model; based on data center server consumption determine the energy demand of the data center; store electricity in the data center energy storage device or obtain electricity from the smart grid; determine the average response time of various types of clients based on the queuing network model; based on energy with the goal of minimizing operating costs The dynamic optimization scheduling model obtains the optimal energy demand configuration of the data center; the intraday timing prediction results of each decision variable in the energy dynamic optimization dispatching model are obtained based on the exponential smoothing method, and the optimal energy within the day of the data center is obtained based on the intraday timing prediction results Requirements configuration. The invention can perform reasonable resource scheduling and energy management on the data center to ensure economical operation.

Description

Energy optimized scheduling method, system and storage medium for data center

Technical field

The present invention relates to the technical field of energy management, and in particular, to an energy optimization scheduling method and system for a data center.

Background technique

In order to adapt to new service needs, data centers continue to improve the performance of servers. At the same time, their heat production is also increasing day by day, and the power of various equipment is showing an upward trend. The resulting energy crisis and greenhouse gas emission issues are also increasingly prominent. At present, renewable energy such as photovoltaics has become the main direction of global energy development in the future, mainly reducing carbon emissions and strengthening environmental protection, cutting energy consumption costs to improve economic benefits, getting rid of excessive dependence on fossil energy on cooling facilities to improve the stability of data center systems, etc. .

The Chinese patent with application number 202011102939.0 and the invention title "A comprehensive energy system optimization dispatching method that integrates distributed data center computing power and energy flow" provides an integrated energy system that integrates distributed data center computing power and energy flow. Optimize the scheduling method, which includes the following steps: establishing a comprehensive energy system mathematical model integrating distributed data center computing power and energy flow based on graph theory; establishing three first-level indicators including economy, safety, and cleanliness and several A comprehensive energy system operation evaluation index system with secondary indicators; a comprehensive evaluation method is used to determine the comprehensive weight of each indicator; the lowest operating cost, the highest safety, and the lowest pollution emissions are used as three objective functions to build an integrated energy system optimal dispatch model; The integrated energy system optimal dispatch model is connected to its mathematical model to obtain the optimal dispatch method and results. This patent incorporates distributed data centers into a comprehensive energy system, and comprehensively optimizes and dispatches computing power, electricity, heat and other energy sources as a whole, which not only reduces energy waste, but also increases the consumption of clean energy and improves system flexibility. sex. However, this patent does not take into account the intermittency, randomness, mutation and other characteristics of clean energy, which may lead to poor economic benefits. Moreover, this patent focuses on the optimal dispatch of the energy supply system and lacks management of the data center energy consumption system.

The Chinese patent application with application number 202110905770.0 and the invention title "A data center real-time energy management method and system" discloses a data center real-time energy management method and system, which belongs to the field of power management. The energy management method includes: establishing data The real-time energy management model of the center in the current preset period, and reconstruct the model into a Markov decision process; optimize the optimal real-time energy management strategy of the data center by solving the Bellman equation period by period, using the The three-dimensional state variables of batch processing load, power storage and cold storage are approximated by value functions respectively to overcome the difficulty of solving the value function. The batch processing load approximate value function, power storage approximate value function and cold storage capacity in the queue obtained by offline training are The approximation function is combined to obtain a three-dimensional approximation function and substituted into the Bellman equation in the Markov decision-making process. The Bellman equation is solved to obtain an approximately globally optimal set of decision variables for managing the data center. This patent application enables real-time energy management of data centers operating in uncertain environments. The problem with this patent application is that it still does not take into account the intermittency, randomness, mutation and other characteristics of clean energy in the energy system, which may lead to poor service stability, and only considers the economics of the data center in terms of energy consumption costs.

The Chinese patent application with application number 202080003421.3 and the invention title "Data Center Energy Management System" discloses technology that includes managing energy flow within the system. The patent application describes a system that includes: a power generation system; having a charging state properties of a battery storage system; and processing circuitry capable of accessing the power grid, generation system, and battery storage system. In one example, the processing circuit is configured to: determine an energy utilization forecast for the data center; monitor energy availability factors; determine a defined energy utilization factor involving the power grid, the power generation system, the battery storage system, and the data center based on the energy utilization forecast and the monitored energy availability factors. An energy flow configuration of the energy flow, wherein the energy flow configuration includes information identifying one or more of a power grid, a power generation system, or a battery storage system as a power source for the data center; providing power to the data center based on the energy flow configuration; and based on the energy flow Configuration management involves the energy flow of battery storage systems. The patent application considers various factors affecting the energy generation system and performs system management based on the energy flow within the system, but does not consider the factors affecting the cost-effectiveness of the data center.

The data center is an energy system with a large amount of new energy access, and it is also a huge physical information system. On the premise of ensuring the economical operation of the data center, it is reasonable to meet the planning and arrangement of computing tasks and to control the energy consumption of the data center. Effective management is crucial to the operation of the entire data center. Proper resource scheduling and energy management are key to minimizing operating costs and ensuring the safety and stability of the data center's operations.

Contents of the invention

In view of this, embodiments of the present invention provide an energy optimization scheduling method for a data center to solve resource scheduling and energy management problems on the energy supply side and energy demand side of the data center.

One aspect of the present invention provides an energy optimization scheduling method for a data center, which method includes the following steps:

Obtain numerical weather prediction data, input the obtained numerical weather prediction data into the pre-trained BiLSTM prediction network model, and output the day-ahead prediction results of photovoltaic power generation output power;

Determine the energy demand of the data center based on the power consumed by the data center servers when starting up, working and cooling;

Determine the amount of electricity to be stored in the data center energy storage device or the amount of electricity to be obtained from the smart grid based on the day-ahead prediction results of the photovoltaic power generation output power and the energy demand of the data center;

Determine the average response time of various types of clients based on the queuing network model;

The optimal energy demand configuration of the data center is obtained based on the energy dynamic optimization scheduling model with the goal of minimizing operating costs; where the decision variables in the energy dynamic optimization scheduling model include at least one of the following variables: arrival of client request rate, real-time electricity price of smart grid and photovoltaic power generation output power; the operating cost is obtained based on the following factors: the amount of electricity exchanged between the data center and the smart grid, the operation and maintenance price of energy storage devices and the average response time of various types of clients;

Based on the exponential smoothing method, the intraday time series prediction results of each decision variable in the energy dynamic optimization scheduling model are obtained. Based on the intraday time series prediction results, the optimal energy demand configuration of the data center within the day is obtained, and the energy dynamic optimization scheduling model is updated according to the time series progression. .

In some embodiments of the present invention, before pre-training the BiLSTM prediction network model, the method further includes:

Raw photovoltaic power output and numerical weather prediction data are processed using standard score normalization.

In some embodiments of the present invention, the power consumption for cooling is determined based on the external air cooling method.

In some embodiments of the present invention, determining the amount of electricity to be stored in the data center energy storage device or the amount of electricity to be obtained from the smart grid based on the day-ahead prediction result of the photovoltaic power generation output power and the energy demand of the data center includes:

If the day-ahead prediction result of photovoltaic power generation output power is less than the energy demand of the data center, the data center obtains electricity from the smart grid, and the energy demand of the data center is equal to the sum of the day-ahead prediction result of photovoltaic power generation output power and the electricity amount obtained by the smart grid; If the day-ahead prediction result of photovoltaic power generation output power is greater than the energy demand of the data center, the energy storage device of the data center stores excess electricity, and the energy demand of the data center is equal to the day-ahead prediction result of photovoltaic power generation output power and the energy storage device storage power. Difference.

In some embodiments of the present invention, the queuing network model is established based on the exponential distribution of customer request arrival time and service time.

In some embodiments of the present invention, the arrival rate of the client request is based on the site effective arrival rate of the client request determined by the routing matrix.

In some embodiments of the invention, the average response time is determined based on Little's law.

In some embodiments of the present invention, the exponential smoothing method includes a seasonal processing method in the Holt-Winters three-parameter exponential smoothing method.

Another aspect of the present invention provides an energy optimization scheduling system for a data center. The system includes a processor and a memory. Computer instructions are stored in the memory. The processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described in any of the above embodiments.

Another aspect of the present invention provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method described in any of the above embodiments are implemented.

The energy optimization dispatching method and system of the data center of the present invention studies the multi-channel collaborative energy of the data center by combining IT equipment and cooling equipment such as renewable photovoltaic energy on the energy supply side, energy storage devices, smart grids, and energy demand side servers. management issues. The advantage of this invention is that the energy application scenario studied is clean and requires multi-channel coordination, which is in line with the current sustainable low-carbon complex data center environment; secondly, the photovoltaic power generation output power is predicted based on the BiLSTM time series future multi-step prediction method , can improve the accuracy of day-ahead prediction results of photovoltaic power generation output power; in addition, the exponential smoothing method is introduced to perform intra-day prediction of each decision variable in the energy dynamic optimization dispatch model with the goal of minimizing operating costs, and continuously predict and optimize based on actual changes , update the optimal energy consumption configuration of the data center, coordinate the output of multi-channel equipment in the system, effectively manage the energy consumption of the data center on the premise of meeting the planning and arrangement of computing tasks, and reasonably solve resource scheduling and energy management implementation Minimize data center operating costs and ensure safe and stable operation.

Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following, or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the specification and drawings.

Those skilled in the art will understand that the objectives and advantages that can be achieved with the present invention are not limited to the specific description above, and the above and other objectives that can be achieved with the present invention will be more clearly understood from the following detailed description.

Description of drawings

The drawings described here are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.

Figure 1 is a schematic flowchart of an energy optimization scheduling method for a data center in an embodiment of the present invention.

Figure 2 is a bar chart of the number of servers allocated to network groups, application groups and database groups in one day in an embodiment of the present invention.

Figure 3 is a bar chart of the number of servers turned on and off in one day according to an embodiment of the present invention.

Figure 4 is a line chart comparing the intra-day predicted value and the actual value of renewable energy production in one day in an embodiment of the present invention.

Figure 5 is a line chart of the power the data center obtains from the smart grid in one day according to an embodiment of the present invention.

Figure 6 is a line chart of the production volume and sales volume of renewable energy in one day in an embodiment of the present invention.

Figure 7 is a line chart of energy sold and used by the energy storage device in one day according to an embodiment of the present invention.

Figure 8 is a line chart comparing the intra-day predicted value and the actual value of the renewable energy sales price in one day in an embodiment of the present invention.

Figure 9 is a line chart comparing the intra-day predicted value and the actual value of the power grid price in one day in an embodiment of the present invention.

Figure 10 is a comparative line chart of data center operating costs obtained through the simulation method and the optimization solution of the present invention in one embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below in conjunction with the embodiments and drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not used to limit the present invention.

Here, it should also be noted that, in order to avoid obscuring the present invention with unnecessary details, only the structures and/or processing steps closely related to the solution according to the present invention are shown in the drawings, and the details related to them are omitted. Other details are less relevant to the invention.

It should be emphasized that the term "comprising" when used herein refers to the presence of features, elements, steps or components but does not exclude the presence or addition of one or more other features, elements, steps or components.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

The data center is an energy system with a large amount of new energy sources. It is also a huge physical information system. Whether it can effectively manage the energy consumption of the data center while meeting the planning and arrangement of computing tasks is critical to the operation of the entire data center. It is crucial that reasonable resource scheduling and energy management are the key for data centers to minimize operating costs and ensure the safety and stability of their operations.

Based on this, the present invention proposes a data center energy management optimization scheduling method including day-ahead and intra-day prediction. First of all, considering the intermittent characteristics of clean energy, the future multi-step energy prediction method based on BiLSTM time series improves the accuracy of day-ahead photovoltaic power generation output power prediction; in addition, considering the energy supply of energy storage devices and smart grid interaction, the server meets multi-application services Factors such as energy consumption of requests and cooling are used to build an energy dynamic optimization scheduling model with the goal of minimizing operating costs; the exponential smoothing method is introduced to predict each decision variable in the model within a day, and the optimal data center is continuously predicted, optimized, and updated based on actual changes. Optimize energy consumption configuration, coordinate the output of multi-channel equipment in the system, reduce the operating cost of the data center, and reduce the adverse effects of difficult predictions and equipment power outages caused by uncertain factors such as weather, power market, and user behavior.

Figure 1 is a schematic flowchart of an energy optimization scheduling method for a data center according to an embodiment of the present invention. As shown in Figure 1, the method may specifically include steps S110 to S160.

Step S110: Obtain numerical weather prediction data, input the obtained numerical weather prediction data into the pre-trained BiLSTM prediction network model, and output the day-ahead prediction result of photovoltaic power generation output power.

Photovoltaic power generation itself has scattered physical locations, small production scale, and power output with characteristics such as intermittent, randomness, and mutation. Therefore, the output power of photovoltaic power generation is often affected by multiple factors such as time and weather, ignoring other meteorological factors. The influence of factors can easily lead to prediction errors. In addition, the accuracy of photovoltaic power generation output power prediction increases the cost of photovoltaic power generation output power prediction.

Numerical Weather Prediction (NWP) refers to a method that uses the mathematical model of the atmosphere and current weather conditions as input data to predict the atmospheric motion status and weather phenomena for a certain period in the future. It is usually completed using supercomputers or distributed computing clusters. Calculations, including meteorological information such as irradiance, wind speed and direction.

More specifically, the pre-trained BiLSTM prediction network model can predict the response of the data and update the network status, and then predict the photovoltaic power generation output power in a certain time period in the future based on the set sliding window range.

Furthermore, the Long Short Term Memory (LSTM) network has a wide range of applications in future multi-step prediction. The LSTM model controls the discarding or adding of information through mechanisms such as forgetting gates, input gates, and output gates, effectively solving the problem of long-term dependence, and is suitable for classification, processing, and prediction based on time series data. However, the LSTM model can only predict the information of the unit at the next moment through the temporal information of the unit at the previous moment, and cannot encode the information from back to front. The Bi-directional Long Short Term Memory (BiLSTM) network is composed of It is composed of forward LSTM and backward LSTM, which can fully learn the information of contextual data.

More specifically, the core idea of the LSTM model is as follows:

1) The forgetting gate f can be used to determine the information discarded from the unit state and is used to control the degree to which the previous unit is forgotten. The input _xt of the current unit and the output ht _-1 of the previous unit selectively filter the information through the sigmoid function. The output value of the sigmoid function is between [0,1], which can selectively pass the information. 0 means completely discarded and 1 means completely passed.

f _t =Sigmoid(W _f [h _t-1 ,x _t ]+b _f );

In the formula, f _t represents the output of the current forgetting gate, and W _f and b _f represent the weight and bias of the forgetting gate respectively.

2) The input gate and tanh function determine some of the new information added to the unit state. The sigmoid function can be used to determine the retained input information i _t , and the x _t and h _t-1 parts are updated by the tanh function to generate a new candidate value C′ _t . Combine i _t and C′ _t to update the old unit state C _t-1 to obtain the new unit state C _t .

i _t =Sigmoid(W _i [h _t-1 ,x _t ]+b _i ];

C _t '=tanh(W _C [h _t-1 ,x _t ]+b _C );

C _t = f _t C _t-1 +i _t ·C'_t;

In the formula, W _i and b _i represent the weight and bias of the input gate respectively, W _c and b _c represent the weight and bias of the tanh layer respectively.

3) The output gate is used to control the degree to which the current unit state is filtered. First, x _t and h _t-1 determine the output part O _t of the unit state through the sigmoid function. Then the unit state is set to tanh and multiplied by O _t to obtain the predicted value h _t of the LSTM model.

O _t =Sigmoid(W ₀ [h _t-1 ,x _t ]+b ₀ );

h _t =O _t ·tanh(C _t );

In the formula, O _t represents the output of the current output gate, W ₀ and b ₀ represent the weight and bias of the output gate respectively.

The single-layer BiLSTM model consists of a forward propagation layer (forward LSTM layer) and a backward propagation layer (reverse LSTM layer), and the forward propagation layer and the backward propagation layer are jointly connected to the output layer. The input sequence is calculated forward from time t ₁ to time t ₂ in the forward propagation layer, and the output of the forward hidden layer at each time is obtained and saved; the input sequence is calculated in the reverse direction from time t ₂ to time t ₁ in the back propagation layer. Calculate once, obtain and save the output of the backward hidden layer at each moment; finally, at each moment between [t ₁ , t ₂ ], combine the output results of the forward propagation layer and the back propagation layer at the corresponding moment to get the final Output. Since the BiLSTM model can better capture bidirectional semantic dependencies, it can predict photovoltaic power generation output power more accurately in the future multi-step day-ahead.

As an example, the specific steps to establish the pre-trained BiLSTM prediction network model may include:

1) Construction of BiLSTM prediction network model. The BiLSTM network model includes a sequence input layer, a BiLSTM layer, a fully connected layer and an output layer, and configures related training parameters such as network weights, learning rate, and number of iterations.

2) Training of BiLSTM prediction network model.

The standardized training data set containing K sets of data Enter the BiLSTM prediction network model constructed in the above steps, where x _k is an independent variable data set, including related factors that affect photovoltaic power generation output power, such as NWP data; y _k is a dependent variable data set, which contains K photovoltaic power generation output powers. Data collection. After dividing the training set and test set samples, the corresponding photovoltaic power generation prediction model is obtained through multiple iterative training and corrections, that is, the pre-trained BiLSTM prediction network model.

Using NWP data as an independent variable to predict photovoltaic power generation output power is only an example, but the present invention is not limited to this. Solar cell components, dust loss and other factors that affect photovoltaic power generation output power can also be used.

In some embodiments of the present invention, before pre-training the BiLSTM prediction network model, it also includes: using standard score normalization to process the original photovoltaic power generation output power and numerical weather prediction data.

More specifically, in order to make the data have the same measurement scale and at the same time the neural network can converge quickly, the standard score (Z-score) method can be used to standardize the input raw NWP data and photovoltaic power generation. The formula of Z-score normalization method can be expressed as:

In the formula, x _i represents the i-th data in the original NWP data and photovoltaic power generation data sample, x′ _i represents the i-th data in the standardized data sample, mean(x) represents the mean of the overall data sample, std(x) represents the standard deviation of the population data sample.

In addition, the output power of photovoltaic power generation needs to be limited as follows:

_0≤Ps (t) _≤Psmax ;

Among them, t∈{0,1,...,T-1}, T represents the predicted time range (can be divided into T-1 time intervals), P _s (t) represents the photovoltaic power station in the t-th time interval The power generation output power, P _smax represents the maximum power generation output power of the photovoltaic power station in the data center.

A specific example is listed below to describe the accuracy of BiLSTM's future multi-step prediction of photovoltaic power output power, including:

In the day-ahead photovoltaic prediction stage, simulation comparison tests were conducted based on BiLSTM single-step prediction, LSTM single-step prediction, BP prediction and BiLSTM multi-step prediction methods. Table 1 shows the prediction evaluation index results of the four simulation comparison methods. Data from January 2017 to August 2018 (data collected every 15 minutes, a total of 65,760 time series related data) are selected to predict photovoltaic power generation data from October to December 2018. By analyzing the day-ahead prediction results of renewable photovoltaic power generation output power, it can be concluded that the BiLSTM multi-step prediction method can achieve higher prediction accuracy.

The training processes between neural network models are mostly similar, but the difference lies in the design of the network structure and the adjustment of parameters. The difference between the LSTM single-step prediction, BiLSTM single-step prediction and BP prediction methods and the BiLSTM multi-step prediction method is that the prediction window is set to 1 in the BiLSTM single-step prediction model; the prediction window is set to 1 in the LSTM single-step prediction model. The BiLSTM layer in the neural network model is replaced by the LSTM layer; the neural network of the BP prediction model is designed as an input layer, a hidden layer, and an output layer. The above four neural network models all require four steps of data preprocessing, model construction, training and prediction.

In addition, the data used in the prediction process include desensitized environmental data (that is, environmental data obtained after removing data with large deviations) and the actual irradiance of the electric field and the power generated by the electric field. The data fields include time, irradiance, and wind speed. , wind direction, temperature, humidity, pressure, actual power.

This invention selects two conventional prediction evaluation indicators: mean absolute error (MAE, Mean Absolute Error) and mean square error (RMSE, RootMean Square Error), both of which range from [0, +∞). When the prediction of photovoltaic power generation output power When the value is completely consistent with the true value, it is equal to 0, which is a perfect model. The greater the error, the greater the value of MAE and RMSE. The evaluation index results of the four prediction methods are shown in the table. It can be seen that the BiLSTM multi-step prediction has the smallest MAE and RSME effects among the four methods, indicating the gap between the photovoltaic power generation value predicted by this method and the real power generation value. The smallest, that is, the BiLSTM multi-step prediction method is the best among the four prediction methods of photovoltaic power generation output power, and the prediction accuracy is higher.

Table 1 Predictive evaluation index results of the four methods

Indicators\Forecasting Methods BiLSTM multi-step BiLSTM single step LSTM single step BP MAE 1.1221 1.4216 1.7633 3.0305 RSME 2.0010 2.3520 2.7385 4.8421

Step S120: Determine the energy demand of the data center based on the power consumed by starting the data center server, the power consumed by working, and the power consumed by cooling.

The energy demand E _total (t) of the data center energy consumption system mainly depends on the energy G _on (t) consumed by starting a new server in the tth time interval, and the energy consumed by the server in the working state E _c (t) , and the energy consumed for server cooling E _cool (t).

As an example, the basic idea of calculating the energy demand of a data center is as follows:

1. Turn on the determination of server power consumption, which may include the following steps:

The increase in the number of servers working in the data center will increase energy consumption. If the energy consumed by each server when it is turned on is E _on (t), then the energy consumed by turning on a new server within the interval t (G _on 9t) can be expressed as:

G _on (t)=N _on (t)E _on (t);

In the formula, N _on (t) represents the number of working servers added within the interval t.

2. Determining the power consumption of servers in working status in the data center may include the following steps:

1) The server's opening and closing status needs to be captured in each continuous time interval. The number of servers working in the interval t is equal to the number of servers in the working state N _w (t-1) in the previous time interval t-1 plus the number of servers turned on and off in the time interval t N _on (t) And subtract the number of turned-on servers that are turned off, N _off (t). The number of servers N _w (t) working within the interval t can be expressed as:

N _w (t)=N _w (t-1)+N _on (t)-N _off (t); or

N _w (t)=N _{w, web} (t) + N _{w, app} (t) + N _{w, db} (t);

Among them, N _w,web (t) is the number of servers working in the network group within the time interval t, N _w,app (t) is the number of servers working in the application group within the time interval t, and N _w,db (t) is The number of servers working in the database group within this time interval t. When t=1, N _w (t-1) is a given constant.

2) If the energy consumed by each server working within a time interval is expressed as E _w (t), then the power consumed by the server working within the interval t E _c (t) can be expressed as:

E _c (t) = N _w (t) E _w = (N _web (t) + N _app (t) + N _db (t)) E _w ;

3. Determining the power consumption of server cooling may include the following steps:

The energy consumption of cooling servers and other IT equipment is directly related to the operating energy consumption of these IT equipment. If IT equipment uses different types of cooling processes, energy consumption calculation methods are also different.

In some embodiments of the present invention, the power consumption for cooling is determined based on the external air cooling method.

The external air cooling method is an air cooling technology. This method does not require expensive coolers. The cooling equipment uses the temperature difference between the inside and outside of the data center IT equipment to use the lower-temperature outside air as a cooling air source and send it to the IT equipment for heat exchange to achieve cooling purposes.

More specifically, the energy consumption E _cool (t) of cooling the server using external air cooling method within the interval t can be expressed as:

In the formula, α represents the relative temperature ratio between the internal and external settings of the IT equipment, b represents the calculated ratio of the internal and external temperatures of the IT equipment, t _in represents the internal temperature of the IT equipment, assuming it is 35°C, and t _out represents the IT equipment. The external temperature depends on the geographical location of the data center.

Using the external air method to determine the power consumption for cooling the data center server is only an example, but the present invention is not limited thereto, and also includes using other cooling methods to calculate the power consumption for cooling the data center server.

More specifically, the calculation formula of the energy demand E _total (t) of the data center within the interval t can be expressed as:

E _total (t) = E _c (t) + E _cool (t) + G _on (t).

Step S130: Determine the amount of electricity to be stored in the data center energy storage device or the amount of electricity to be obtained from the smart grid based on the day-ahead prediction result of the photovoltaic power generation output power and the energy demand of the data center.

Renewable photovoltaic energy, energy storage devices and smart grids together constitute the energy system.

Due to the continuous expansion of renewable energy application fields and the acceleration of the openness of the power market, most of the current energy systems of data centers have energy exchanges with the external power grid. Therefore, in order to ensure the energy supply and energy consumption of the energy system, Energy demand is in a balanced state. When the output power of renewable photovoltaic power generation is less than the energy consumption of the data center, the external power grid supplies power to the data center; when the output power of renewable photovoltaic power generation is greater than the energy consumption of the data center, the data center transmits power to the external power grid. .

In some embodiments of the present invention, the amount of electricity stored in the data center energy storage device or the amount of electricity that needs to be obtained from the smart grid is determined based on the day-ahead prediction result of the photovoltaic power generation output power and the energy demand of the data center, including:

If the day-ahead prediction result of photovoltaic power generation output power is less than the energy demand of the data center, the data center obtains electricity from the smart grid, and the energy demand of the data center is equal to the sum of the day-ahead prediction result of photovoltaic power generation output power and the electricity amount obtained by the smart grid; If the day-ahead prediction result of photovoltaic power generation output power is greater than the energy demand of the data center, the energy storage device of the data center stores excess electricity, and the energy demand of the data center is equal to the day-ahead prediction result of photovoltaic power generation output power and the energy storage device storage power. Difference.

More specifically, the basic idea that the energy supply of the energy system and the energy demand of the energy consumption system are in balance is as follows:

1) If the day-ahead prediction result of photovoltaic power generation output power is greater than the energy demand of the data center, the energy storage device of the data center needs to be used to store excess energy.

In the process of transferring excess power from the photovoltaic power station in the data center to the energy storage device, the operating variables of the energy storage device need to be constrained. The constraint variables of the energy storage device mainly consider the possible maximum and minimum capacities of the energy storage device, charging power and discharging power, charging efficiency and discharging efficiency, and dynamic changes in the amount of electricity stored in the energy storage device. Variables that affect the operation of energy storage devices can be limited as follows:

In the formula, Respectively represent the minimum and maximum charging power, minimum and maximum discharge power, minimum and maximum capacity levels of the energy storage device,/> Respectively represent the charging power and discharge power of the energy storage device in period t (also called the tth time interval or interval t), μ _c (t), μ _d (t) respectively represent the charging state and discharge state of the energy storage device in period t , is a 0/1 variable. E _esd (t) represents the capacity of the energy storage device in period t. When t=0, the capacity of the energy storage device is E _esd , which is generally a constant. /> Indicates the state of transition from charging to discharging,/> Indicates the state of transition from discharge to charge, and is a 0/1 variable. C _esd represents the upper limit of the number of charging and discharging times of the energy storage device in a day, α _esd and β _esd are given percentages, and/> represents the charging efficiency and discharge efficiency of the energy storage device in the data center, δ _esd represents the self-discharge efficiency of the energy storage device, and Δt is the time interval difference.

in, and/> is defined as follows:

2) If the day-ahead prediction result of photovoltaic power generation output power is less than the energy demand of the data center, in order to meet the energy demand of the data center and make the server operate normally, the data center needs to obtain power from the smart grid.

The variables involved in the process of exchanging energy between smart grids and data centers need to be limited as follows:

In the formula, Respectively represent the minimum and maximum power transmitted from the smart grid to the data center and the minimum and maximum power transmitted from the data center to the smart grid. In actual circumstances, it is generally believed that/> P _b (t) represents the power transmitted from the smart grid to the data center in period t, P _se (t) represents the power transmitted from the data center to the smart grid in period t, μ _b (t) represents the power transmitted from the smart grid to the data center in period t The state of , μ _se (t) respectively represents the state of power transmission from the data center to the smart grid during t period, and both are 0/1 variables.

Among them, the variables μ _b (t), μ _se (t) are defined as follows:

3) If the day-ahead prediction result of photovoltaic power generation output power is equal to the energy demand of the data center, there is no need to store power in the data center energy storage device or obtain power from the smart grid.

If the day-ahead prediction result of photovoltaic power generation output power is equal to the energy demand of the data center, the generated renewable energy will just meet the energy demand of IT equipment such as servers. The data center does not need to obtain power from the smart grid, and there will be no excess power delivered to storage device for storage.

Step S140: Determine the average response time of various types of clients based on the queuing network model.

The service system of the data center includes all servers in the data center and a scheduler dis.

The service system of the data center can be regarded as a queuing network model. Each request sent by the client to the data center is assigned accordingly by the scheduler at the entrance of the data center to the network group web, application group app or database group db, using each Group-assigned servers fulfill requests and the number of servers assigned to each group changes continuously. If a request leaves a group, it may go to another group or leave the data center, depending on the routing probabilities between network groups, application groups, or database groups.

In some embodiments of the present invention, the queuing network model is established based on the exponential distribution of customer request arrival time and service time. Assuming that the arrival time and service time of client requests are exponentially distributed, the scheduler can be modeled as an M/M/1 queue, and the network group, application group, or database group can be modeled as an M/M/k queue. (where k=N _web , N _app or N _db , N _web is the number of servers in the network group, N _app is the number of servers in the application group, N _db is the number of servers in the database group, and the sum of the number of servers in the three groups is data The total number of servers in the center is N). In addition, considering the complex inference of the average response time of each group of multi-server queues, the network group, application group or database group can be modeled as an independent server group with the same arrival rate, and each server within each group can be modeled as An M/M/1 queue.

More specifically, assume that the service system is a Jackson queuing network. The Jackson network is a set of queues with different routing probabilities, and the scheduler, network group, application group and database group queues in the network are also called sites.

In some embodiments of the present invention, the arrival rate of the client request is based on the site effective arrival rate of the client request determined by the routing matrix.

More specifically, the arrival rate of client requests is the effective arrival rate of each site, which can be calculated through the routing matrix P, and the client of each type i (also called customer type i, assuming there are N _c types The client, 1≤i≤N _c ) will have different matrices _Pi , and the formula can be expressed as:

The arrival rate of each site is used as the performance measure of each site to evaluate the performance of the entire network. The vector λ _i contains the effective arrival rate of each site calculated as follows:

In the formula, I represents the identity matrix, Indicates the effective site arrival rate of customer type i.

Since both the scheduler and each server in each group can be modeled as an M/M/1 queue, for client type i, the average number of requests L _i,dis of the scheduler can be calculated as the same as that of each server in each group. The average number of requests L _i,k is calculated in the same way (the average number of requests is also called the queue length), and the formula is expressed as:

In the formula, s _k represents the site transfer probability, ρ _k or ρ _dis is an intermediate calculation variable, λ _i,k is the arrival rate of requests of customer type i to each server in each group, λ _i,dis is the request of customer type i Arrival rate to the scheduler, μ is the service rate of each server in each group, μ _dis is the service rate of the scheduler, assuming μ _dis >> μ.

For customer type i, the average number of requests in each group is the sum of the average number of requests of the servers in each group, and the sum of the average number of requests at each site is the average number of requests _Li for customer type i in the service system.

L _i =L _i,dis +L _i,web +L _i,app +L _i,db .

In some embodiments of the invention, the average response time is determined based on little's law.

The calculation formula of the average response time can be expressed as:

In terms of data center services, a service level agreement (SLA, Service Level Agreement) is signed between the data center and N _c types of clients. These different types of clients have different arrival behaviors and different request sizes throughout the day. , and the number of servers in each group is also related to the processing of requests, so each type of client has different average response times. Since the average response time can be used as an evaluation indicator of SLA, each customer type i has a cost function M _i that depends on the average response time, and the formula can be expressed as:

If the data center performance can ensure that the average response time W _avg,i of a certain type of client i within the interval t is lower than the critical time W _c (that is, the average response time for client type i specified in the SLA), no loss will occur and the data The center will receive the benefit a _i . However, if the average response time W _avg,ii exceeds the critical time, the data center will have to pay a penalty that increases in proportion to the penalty slope k _i .

The present invention's determination of the average response time of various types of clients based on the Jackson queuing network is only an example, and other types of queuing networks can also be used to determine the average response time.

Step S150: Obtain the optimal energy demand configuration of the data center based on the energy dynamic optimization scheduling model with the goal of minimizing operating costs; wherein the decision variables in the energy dynamic optimization scheduling model include at least one of the following variables: Client The arrival rate of the request, the real-time electricity price of the smart grid and the output power of photovoltaic power generation; the operating cost is obtained based on the following factors: the amount of electricity exchanged between the data center and the smart grid, the operation and maintenance price of the energy storage device and the average response time of various types of clients.

More specifically, based on the photovoltaic power generation output power and the energy demand of the data center, the data center can determine the exchange of energy with the smart grid or the excess energy stored in the energy storage device; and then based on the average response time of different types of clients, the data center can be known The benefits of the SLA between the data center and the client; using the power exchanged between the data center and the smart grid, the dynamic operation of the energy storage device charging/discharging and the SLA benefits as factors affecting the data center operating costs, it is constructed to minimize the operating costs as The target energy dynamic optimization scheduling model can obtain the optimal energy consumption configuration of the data center and determine the number of servers that need to be turned on and off within the interval t. Therefore, the operating status of the IT equipment and cooling equipment in the data center can be known.

Considering the economic benefits of data center operation, the energy dynamic optimization scheduling model with the goal of minimizing operating costs can be expressed as:

In the formula, PR _g is the real-time electricity price when the data center trades with the smart grid, and M _esd is the price of operation and maintenance of the energy storage device. According to the expression of the average response time and the calculation of the energy consumption of the cooling method, it can be seen that the optimization problem is non-linear, and the cost objective function focuses on the cost of the data center in terms of satisfying the client's request and energy consumption, and also includes the situation where the energy can be sold possible benefits. Based on the energy dynamic optimization scheduling model, the optimal energy demand configuration of the data center can be obtained.

The arrival rate of client requests, real-time electricity prices of smart grids and photovoltaic power generation output power can be used as decision variables of the energy dynamic optimization dispatch model.

The arrival rate of client requests, the real-time electricity price of the smart grid and the output power of photovoltaic power generation as decision variables of the energy dynamic optimization dispatch model are only examples. Other indicators such as energy sales prices can also be used as decision variables of the energy dynamic optimization dispatch model.

Step S160: Obtain the intraday time series prediction results of each decision variable in the energy dynamic optimization scheduling model based on the exponential smoothing method, obtain the optimal energy demand configuration of the data center within the day based on the intraday time series prediction results, and update the energy dynamics according to the time series progression. Optimize scheduling model.

More specifically, during the actual operation of the data center, it is necessary to predict the decision variables of the energy dynamic optimization scheduling model such as the arrival rate of client requests at interval t, the real-time electricity price of the smart grid, and the output power of renewable photovoltaic power generation. Based on the energy dynamic optimization scheduling model with the goal of minimizing operating costs, the optimal energy consumption configuration of the data center is obtained, and the number of servers that need to be turned on and off within the interval t is determined. Therefore, the operating status of the IT equipment and cooling equipment in the data center can be known. And according to the time sequence progression, the predicted value of the next time interval is replaced with the real value, and the operating status of the data center equipment is continuously updated.

Time series prediction needs to predict several parameters when solving optimization problems, because the actual values of these parameters continue to change over time and can only be known within the corresponding time interval. However, in order to ensure energy supply and energy consumption balance and data center operation To minimize costs, planning decisions must be made in advance. The parameters that need to be predicted include the arrival rate of client requests, real-time electricity prices of smart grids, and renewable photovoltaic power generation output power.

Exponential smoothing is a widely used forecasting technique. This method assigns weights to different observations of the time series Y _t , where newer observations receive more weight than older observations. When the time series exhibits seasonal behavior (seasonal cycle with duration s), this method can make predictions by using the seasonal characteristics of the series.

In some embodiments of the present invention, the exponential smoothing method includes a seasonal processing method in the Holt-Winters three-parameter exponential smoothing method.

Since renewable photovoltaic power generation exhibits seasonal behavior, the seasonal processing method in the Holt-Winters three-parameter exponential smoothing method can complete the prediction process of each decision variable in the energy dynamic optimization dispatch model, which requires the use of two parameters: L _t and S _t , its calculation formula can be expressed as:

In the formula, 0γ, κ≤1, κ is the horizontal smoothing coefficient, γ is the seasonal smoothing coefficient, s is the total time, L _t represents the horizontal estimated value within the interval t, Y _t is the actual value of the time series within the interval t, S _t is the seasonal estimate within the interval t, and S _ts is the seasonal estimate within the interval ts.

In the next time interval t+1, the predicted value F _t+1 of the time series can be expressed as:

F _t+1 =L _t *S _t-s+1 ;

The first t values of S and the initialization method of L _t can be expressed as:

More specifically, the process of predicting the decision variables of the energy dynamic optimization dispatch model within a day using the exponential smoothing method is:

First, the two parameters L _t and S _t are initialized according to the above formula, and then the predicted value F t+ ₁ of the next time interval t+1 is obtained according to the exponential smoothing formula, and the predicted value F _t+1 is added to the constructed energy In the dynamic optimization scheduling model, the service resource allocation and energy consumption configuration plan for the interval t is obtained through Matlab. When time t+1 is reached, the F _t+1 predicted in the previous interval is updated and replaced with the actual data Y t ₊ 1 of the new time interval t+1, and then the actual data of the current time series, namely Y ₁ , Y ₂ ,…,Y _t ,Y _t+1 , and then calculate the parameters L _t+1 , S _t+1 , F _t+2 according to the formula, and loop the above process until the end of time.

Another specific embodiment is listed below to compare the predicted data with the actual operation of the data center, which mainly includes the following content:

In this specific embodiment, one day is selected as the prediction time unit. Compared with long-term prediction, the prediction result is more accurate. Tests are conducted at 1-hour intervals. Each interval requires forecasting of energy prices, renewable energy generation and customer arrival rates. These forecast data are used to execute the optimization scheme proposed by the present invention (at intervals) to determine the allocation to each customer. The number of servers in each group and the values of other decision variables.

1) Dynamic operation of data center service system:

Figure 2 depicts the changes in the number of servers assigned to each group during the day, and the trend of these number changes is similar to the trend of the objective function. As can be seen in Figure 2, the server assigns the most to the App type (application software), followed by the Database type (database), and the least to the Web type (website). This is in line with the fact that in real life, the number of service requests on the mobile phone is larger.

Figure 3 depicts the changes in the number of servers turned on and off during the day, and the trends in these numbers are similar to those of the objective function. As can be seen from Figure 3, there is no situation where all servers are turned on at any hour of the day, which can reduce costs.

2) The situation of power exchange between data center energy supply system and smart grid:

Figure 4 is a comparison between the true value of renewable energy production in a day and its predicted value within the day. It can be seen from Figure 4 that there is a certain error between the true value and the predicted value, but the current average error range is -1 %~1%, the error range is small.

Figure 5 shows the energy purchased by the data center from the smart grid during a day. As can be seen from Figure 5, due to the strong radiation during the 10:00-14:00 time period, enough renewable photovoltaic energy is generated, and the data center does not need to purchase power from the smart grid. Other time periods such as 4:00-8:00 Due to the limited photovoltaic power generation capacity during the 16:00-20:00 time period, it cannot meet the energy consumption needs of the data center, and more power needs to be purchased from the smart grid.

Figure 6 shows the changes in production and sales of renewable photovoltaic energy within a day. It can be seen from Figure 6 that there is sufficient sunshine between 7:00 and 18:00. The closer to noon, the greater the radiation intensity and the higher the renewable photovoltaic power generation output power. At other times there is no radiation and little renewable energy is produced. It can be seen from Figure 5 that at other times of the day except the 10:00-14:00 time period, the data center needs to purchase power from the power grid, which shows that the production volume of renewable photovoltaic energy in other time periods does not meet the data. Center energy requirements.

Figure 7 shows the stored energy sales and capacity usage of the energy storage device within a day. If a certain number of servers are shut down, the data center gets leftover power generation that can be sold. The energy storage device provides energy at all times to ensure stable and reliable operation of the system. According to the renewable photovoltaic power generation situation and server energy consumption, part of the stored electric energy is sold during the 11:00-14:00 time period. This is because the sales price is greater than the sales price of renewable energy and can generate more benefits.

Figure 8 depicts the difference between intraday forecast results and actual values of renewable energy sales prices within a day. There are small deviations from the true value in several time periods, and the average error is -0.01%, which is a small error. It can be seen from the results that the performance of the prediction method is good enough to obtain good prediction data.

Figure 9 depicts the difference between the intraday forecast results and actual values of smart grid energy costs within a day, and the forecast trend is roughly the same as Figure 8. It can be seen from Figure 9 that there is a small deviation between the real value and the actual value in several time periods, and the average error is -0.01%, which is a small error. It can be seen from the results that the performance of the prediction method is good enough to obtain good prediction data.

3) The prediction effect of the optimization plan of the present invention:

Figure 10 shows the target cost comparison results between the simulation method and the optimization scheme of the present invention. The simulation method is randomly scheduled by the data center and does not consider the cost of the data center management solution. There is a difference between the results of the optimization scheme of the present invention and the simulation results, and the optimization scheme is far better than the simulation method. The simplified calculation of the average response time is always an upper bound on the actual average response time given by M/M/k queues. The main assumption of the simplified model is that servers belonging to a group distribute their workload equally and work independently. This is unlike the case with M/M/k queues, where the system dynamically distributes the workload among servers so that they utilize resources in a better way.

In summary, it can be seen that the optimization method that combines pre- and post-day predictions is better than the simulation method, and the intra-day predictions of various decision variables are not much different from the real data. It can not only guarantee a certain service agreement, but also achieve the goal of reducing costs and reducing costs. The purpose of energy consumption.

The data center energy optimization scheduling method proposed by this invention includes day-ahead and intra-day predictions. It uses a large amount of historical data to build a dynamic nonlinear optimization model with the goal of minimizing operating costs based on day-ahead predictions, and more accurately determines the data center at a certain date through intra-day predictions. The advantages of optimal energy configuration for timing are:

1. Aiming at optimizing the scheduling problem of complex data centers in uncertain environments. In terms of energy supply, this invention not only considers renewable energy photovoltaic output and energy storage charging and discharging, but also introduces smart grids for energy interaction of real-time electricity prices to make full use of renewable energy; secondly, it also considers external cooling consumption to comprehensively solve many complex problems in actual operation. Road data center optimization scheduling problem.

2. In view of the intermittent and random characteristics of renewable resources, the present invention proposes a future multi-step day-ahead prediction method based on BiLSTM time series to predict photovoltaic output power day-ahead; request arrival rate, power grid price, and availability within days based on exponential smoothing method Decision-making variables such as renewable power generation are used to derive server resource allocation and energy consumption plans based on the constructed optimal scheduling model. Decision-making variables such as the number of servers in each group and energy consumption configuration are multi-predicted. System data center service requests and renewable energy are considered in many aspects. Energy output and smart grid prices change in real time to ensure the safety and economy of data center operations.

3. Due to real-time changes in service requests, the present invention constructs three types of energy system unit models of supply-side photovoltaic power generation, energy storage devices, and smart grid interaction, and energy demand models of demand-side server consumption and cooling consumption, which are jointly constructed by both sides to minimize A dynamic nonlinear optimization model targeting operating costs uses queuing theory to solve service transfers in the network, and then uses intraday exponential smoothing to perform time series predictions. Optimal scheduling is performed based on the results of a certain time series prediction, and the optimal data center at that time series is obtained. Energy configuration, according to the progression of time series, the model is continuously predicted, optimized, and updated. The effectiveness and superiority of the method proposed in this article is verified through result analysis when the data center is actually running.

The invention can effectively manage the energy consumption of the data center on the premise of satisfying the planning and arrangement of computing tasks, and reasonably solve the resource scheduling and energy management, thereby minimizing operating costs and achieving stable operation of the data center.

Corresponding to the above method, the present invention also provides an energy optimization dispatching system for a data center. The energy optimization dispatching system for the data center includes a computer device. The computer device includes a processor and a memory. The memory stores a computer Instructions, the processor is configured to execute computer instructions stored in the memory, and when the computer instructions are executed by the processor, the energy optimization scheduling system of the data center implements the steps of the foregoing method.

Embodiments of the present invention also provide a computer-readable storage medium on which a computer program is stored. The computer program, when executed by a processor, implements the steps of the foregoing edge computing server deployment method. The computer readable storage medium may be a tangible storage medium such as random access memory (RAM), memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

Those of ordinary skill in the art should understand that each exemplary component, system and method described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of both. Whether it is implemented in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each specific application, but such implementations should not be considered to be beyond the scope of the present invention. When implemented in hardware, it may be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, or the like. When implemented in software, elements of the invention are programs or code segments used to perform the required tasks. The program or code segments may be stored in a machine-readable medium or transmitted over a transmission medium or communications link via a data signal carried in a carrier wave.

It is to be understood that this invention is not limited to the specific arrangements and processes described above and illustrated in the drawings. 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 the present invention 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 between steps after understanding the spirit of the present invention.

In the present invention, features described and/or illustrated with respect to one embodiment may be used in the same or in a similar manner in one or more other embodiments and/or may be combined with or substituted for features of other embodiments. Features of Embodiments.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various modifications and changes may be made to the embodiments of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims (10)

1. An energy optimized dispatching method for a data center is characterized by comprising the following steps:

acquiring numerical weather prediction data, inputting the acquired numerical weather prediction data into a pre-trained BiLSTM prediction network model, and outputting a day-ahead prediction result of photovoltaic power generation output power;

determining an energy requirement of the data center based on the power consumed by the data center server for starting, the power consumed by the work and the power consumed by the cooling;

determining the electric quantity stored in an energy storage device of a data center or the electric quantity required to be acquired from a smart grid based on the solar-day prediction result of the photovoltaic power generation output power and the energy demand of the data center;

determining average response times of various types of clients based on the queuing network model;

obtaining an optimal energy demand configuration of the data center based on an energy dynamic optimization scheduling model aiming at minimizing the operation cost; wherein the decision variables in the energy dynamic optimization scheduling model include at least one of the following variables: the arrival rate of the client request, the real-time electricity price of the smart grid and the photovoltaic power generation output power; the operating cost is obtained based on the following factors: the electric quantity exchanged by the data center and the intelligent power grid, the operation and maintenance price of the energy storage device and the average response time of various types of clients;

And obtaining a daily time sequence prediction result of each decision variable in the energy dynamic optimization scheduling model based on an exponential smoothing method, obtaining the optimal energy demand configuration in the data center day based on the daily time sequence prediction result, and progressively updating the energy dynamic optimization scheduling model according to the time sequence.

2. The method of claim 1, further comprising, prior to pre-training the BiLSTM predictive network model:

and (5) processing the original photovoltaic power generation output power and the numerical weather prediction data by using standard fraction normalization.

3. The method of claim 1, wherein the amount of power consumed by cooling is determined based on an outside air cooling method.

4. The method of claim 1, wherein determining the amount of power stored to a data center energy storage device or the amount of power required to be harvested from a smart grid based on the pre-day prediction of the photovoltaic power generation output power and the energy demand of the data center comprises:

if the day-ahead prediction result of the photovoltaic power generation output power is smaller than the energy requirement of the data center, the data center acquires electric quantity from the intelligent power grid, and the energy requirement of the data center is equal to the sum of the day-ahead prediction result of the photovoltaic power generation output power and the electric quantity acquired by the intelligent power grid; if the day-ahead prediction result of the photovoltaic power generation output power is larger than the energy demand of the data center, the energy storage device of the data center stores redundant electric quantity, and the energy demand of the data center is equal to the difference between the day-ahead prediction result of the photovoltaic power generation output power and the electric quantity stored by the energy storage device.

5. The method of claim 1, wherein the queuing network model is built based on an exponential distribution of client request arrival times and service times.

6. The method of claim 1, wherein the arrival rate of the client requests is an effective arrival rate of the sites of the client requests determined based on a routing matrix.

7. The method of claim 1, wherein the average response time is determined based on litter's law.

8. The method of claim 1, wherein the exponential smoothing method comprises seasonal processing in Holt-windows three-parameter exponential smoothing.

9. An energy-optimized dispatching system for a data center, comprising a processor and a memory, wherein the memory has stored therein computer instructions for executing the computer instructions stored in the memory, which system, when executed by the processor, implements the steps of the method of any of claims 1 to 8.

10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the steps of the method according to any one of claims 1 to 8.