Data center energy system-oriented end-to-end prediction optimization model training method

Abstract

The invention discloses a data center energy system-oriented end-to-end prediction optimization model training method and device, a computer readable storage medium, equipment and a computer program product, and belongs to the field of energy system scheduling. Comprising the following steps: establishing a dual optimization training model of a first operation scheduling optimization training model of the data center energy system under a first operation target; according to the boundary condition information corresponding to the first operation scheduling optimization training model, establishing an uncertain condition prediction training model of the data center energy system in a long short-term memory network; determining a subgradient function of a decision loss function of the uncertain condition prediction training model under the dual optimization training model; and iteratively updating parameters of the uncertain condition prediction training model through a subgradient function to obtain a trained end-to-end prediction optimization model. According to the method provided by the embodiment of the invention, the problem that the accuracy and optimality of the scheduling strategy are unbalanced during resource scheduling of the energy system is solved.

Description

Technical Field

[0001] This application belongs to the field of energy system scheduling, and particularly relates to a training method, device, computer-readable storage medium, equipment, and computer program product for an end-to-end prediction optimization model for a data center energy system. Background Art

[0002] For modern data centers, due to the volatility and uncertainty of the load demand of the data center, the scheduling and optimization of the energy system have become increasingly complex and challenging. In order to ensure the stable operation of the data center energy system and the balance of energy supply and demand, it is necessary to develop efficient scheduling optimization methods so that the data center can achieve optimal resource allocation and scheduling strategies under uncertain conditions.

[0003] With the development of deep learning and artificial intelligence technologies, data-driven two-stage prediction and post-optimization methods have also been widely used. These methods establish a load prediction model through machine learning techniques and make optimization decisions based on the prediction results. The end-to-end prediction optimization method further incorporates the optimization objective into the prediction stage, directly considering the impact of the prediction on the optimization decision, making the entire prediction optimization process form an integrated whole. Common methods used in the above process include stochastic programming or robust optimization.

[0004] However, stochastic programming relies on accurate probability distributions that are difficult to obtain, while robust optimization may lead to overly conservative scheduling strategies, increasing system costs. Summary of the Invention

[0005] This application aims to provide a training method, device, computer-readable storage medium, equipment, and computer program product for an end-to-end prediction optimization model for a data center energy system, which at least solves the problem of the imbalance between the accuracy and optimality of the scheduling strategy during the resource scheduling optimization of the energy system.

[0006] In a first aspect, an embodiment of this application discloses a training method for an end-to-end prediction optimization model for a data center energy system, including: Establish a dual optimization training model of the first operation scheduling optimization training model of the data center energy system under the first operation objective; According to the boundary condition information corresponding to the first operation scheduling optimization training model, establish an uncertain condition prediction training model of the data center energy system in a long short-term memory network; Determine the subgradient function of the decision loss function of the uncertain condition prediction training model under the dual optimization training model; Iteratively update the parameters of the uncertain condition prediction training model through the subgradient function to obtain a trained end-to-end prediction optimization model; the end-to-end prediction optimization model is used to determine the scheduling prediction dual solution of the dual optimization model of the first operation scheduling optimization model according to the boundary condition information corresponding to the first operation scheduling optimization model.

[0007] In a second aspect, an embodiment of the present application also discloses an end-to-end prediction optimization scheduling method for a data center energy system, including: Establish a dual optimization model of the first operation scheduling optimization model of the data center energy system under the first operation objective; Input the boundary condition information corresponding to the first operation scheduling optimization training model into the end-to-end prediction optimization model trained by the training method of the end-to-end prediction optimization model for the data center energy system described in the first aspect to obtain the scheduling prediction dual solution of the dual optimization model; Use the scheduling prediction dual solution as the boundary condition to solve the first operation scheduling optimization model to obtain the scheduling prediction solution of the first operation scheduling optimization model, and use the scheduling prediction solution as the scheduling configuration to perform resource scheduling on the data center energy system.

[0008] In a third aspect, an embodiment of the present application also discloses a training device for an end-to-end prediction optimization model for a data center energy system, including: A training dual construction module for establishing a dual optimization training model of the first operation scheduling optimization training model of the data center energy system under the first operation objective; A training prediction construction module for establishing an uncertain condition prediction training model of the data center energy system in a long short-term memory network according to the boundary condition information corresponding to the first operation scheduling optimization training model; A subgradient construction module for determining the subgradient function of the decision loss function of the uncertain condition prediction training model under the dual optimization training model; A model training module for iteratively updating the parameters of the uncertain condition prediction training model through the subgradient function to obtain a trained end-to-end prediction optimization model; the end-to-end prediction optimization model is used to determine the scheduling prediction dual solution of the dual optimization model of the first operation scheduling optimization model according to the boundary condition information corresponding to the first operation scheduling optimization model.

[0009] In a fourth aspect, an embodiment of the present application also discloses an end-to-end prediction optimization scheduling device for a data center energy system, including: A dual construction module for establishing a dual optimization model of the first operation scheduling optimization model of the data center energy system under the first operation objective; A parameter input module, configured to input boundary condition information corresponding to the first operation scheduling optimization training model into an end-to-end prediction optimization model trained by the training method of the end-to-end prediction optimization model for a data center energy system described in the first aspect, so as to obtain a scheduling prediction dual solution for the dual optimization model. A solution scheduling module, configured to use the scheduling prediction dual solution as a boundary condition to solve the first operation scheduling optimization model, so as to obtain a scheduling prediction solution of the first operation scheduling optimization model, and use the scheduling prediction solution as a scheduling configuration to perform resource scheduling on the data center energy system.

[0010] In a fifth aspect, an embodiment of the present application further discloses a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps described in the first aspect or the second aspect are implemented.

[0011] In a sixth aspect, an embodiment of the present application further discloses an electronic device, including a processor, a memory, and a computer program stored on the memory and executable on the processor, and when the computer program is executed by the processor, the steps described in the first aspect or the second aspect are implemented.

[0012] In a seventh aspect, an embodiment of the present application further discloses a computer program product, on which a computer program is stored, and when the computer program is executed by a processor, the steps described in the first aspect or the second aspect are implemented.

[0013] In summary, in the embodiments of the present application, by constructing a dual optimization training model, the uncertain parameters are transferred from the constraint conditions to the objective function, enabling the optimization model to better capture the impact of the uncertain parameters and reducing the dependence on the exact probability distribution. Furthermore, by establishing an uncertain condition prediction training model in the long short-term memory network, the time series characteristics and volatility of the data center load are accurately captured. Based on the boundary condition information of the dual optimization training model, it is ensured that the uncertain condition prediction training model is consistent with the optimization objective. Then, by determining the subgradient function of the decision loss function, an effective gradient update mechanism is established. Utilizing the characteristic that the subgradient function can handle the optimization problem of non-smooth functions, the training process of the prediction model is made more stable and efficient, reducing the impact of prediction errors on the optimization results and improving the accuracy and robustness of the scheduling decision. Finally, by iteratively updating the parameters of the uncertain condition prediction training model using the subgradient function, an end-to-end integrated training of prediction and optimization is achieved, enabling the prediction results to directly serve the optimization objective, enhancing the overall optimization effect of the model in uncertain scenarios, and ensuring the stable operation of the energy system and the efficient allocation of resources. Thus, based on the method of the embodiments of the present application, a closed-loop optimization training process is formed. Through the training and scheduling of the end-to-end prediction optimization model, the efficient scheduling optimization of the data center energy system under uncertain conditions is realized. Using the dual optimization model and the subgradient update mechanism, the robustness and decision-making accuracy of the model are enhanced, and the problem of the imbalance between the accuracy and optimality of the scheduling strategy during the resource scheduling optimization of the energy system is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] By reading the detailed description of the preferred embodiments below, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present application. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings: Figure 1 is a flowchart of the steps of a method for training an end-to-end prediction optimization model for a data center energy system provided by an embodiment of the present application; Figure 2 is a flowchart of the steps of another method for training an end-to-end prediction optimization model for a data center energy system provided by an embodiment of the present application; Figure 3 is an example of a long short-term memory network provided according to an embodiment of the present application; Figure 4 is a flowchart of the steps of an end-to-end prediction optimization scheduling method for a data center energy system provided by an embodiment of the present application; Figure 5 is the complete process of prediction optimization scheduling under the embodiments of the present application; Figure 6 is the data flow process of predictive optimization scheduling under the embodiments of this application; Figure 7 is a schematic structural diagram of a training device for an end-to-end predictive optimization model for a data center energy system provided by the embodiments of this application; Figure 8 is a schematic structural diagram of an end-to-end predictive optimization scheduling device for a data center energy system provided by the embodiments of this application; Figure 9 is a block diagram of an electronic device provided by the embodiments of this application; Figure 10 is a block diagram of another electronic device provided by the embodiments of this application. Detailed implementation manners

[0015] The exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that this application can be more thoroughly understood and the scope of this application can be fully conveyed to those skilled in the art.

[0016] Figure 1 is a method for training an end-to-end predictive optimization model for a data center energy system provided by this embodiment, specifically including the following steps: Step 101, establish a dual optimization training model of the first operation scheduling optimization training model of the data center energy system under the first operation target.

[0017] In some embodiments of this application, in order to transform the optimization scheduling problem of the data center energy system into a more easily solvable form, a dual optimization training model of the first operation scheduling optimization training model of the data center energy system under the first operation target will be established. The dual optimization training model transfers the uncertain parameters from the constraint conditions to the objective function, enabling the optimization model to better capture the influence of the uncertain parameters. This reduces the dependence on the exact probability distribution and enhances the robustness and adaptability of the model in dealing with uncertainty problems.

[0018] In a specific example, the optimization scheduling of the data center energy system needs to consider the output fluctuations of photovoltaic power generation and the changes in the load demand of the data center. The user first establishes a day-ahead (the next day) scheduling optimization model that includes the output of renewable energy and the load demand. Then, the user performs a dual transformation on this optimization model to form a dual optimization training model. Through the dual optimization training model, the user can effectively capture the influence of these uncertain parameters and reduce the dependence on the exact probability distribution in the subsequent training process.

[0019] Step 102: Establish an uncertain condition prediction training model for the data center energy system in the long short-term memory network according to the boundary condition information corresponding to the first operation scheduling optimization training model.

[0020] In some embodiments of the present application, in order to accurately predict the load demand of the data center energy system and thus improve the accuracy and efficiency of scheduling optimization, an uncertain condition prediction training model for the data center energy system will be established in the long short-term memory (LSTM) network according to the boundary condition information corresponding to the first operation scheduling optimization training model. The LSTM network can capture long-term and short-term dependencies in time series data and is particularly suitable for dealing with load prediction problems with time dependence. By using the LSTM network, the uncertain condition prediction training model can more accurately predict the load demand of the data center, thereby providing reliable input for subsequent optimal scheduling.

[0021] In a specific example, the data center energy system needs to predict the data center power load for the next day. First, according to the historical load data and boundary condition information, input features are prepared, including time-of-use power load data, weather data, and time information. Then, these feature data are input into the LSTM network to train the load prediction model. Through the training of the LSTM network, the model can capture the time series characteristics of the load data and accurately predict the future power load demand.

[0022] Step 103: Determine the subgradient function of the decision loss function of the uncertain condition prediction training model under the dual optimization training model.

[0023] In some embodiments of the present application, in order to effectively update the parameters of the uncertain condition prediction training model during the optimization process and handle the optimization problem of non-smooth functions, the subgradient function of the decision loss function of the uncertain condition prediction training model under the dual optimization training model will be determined. The subgradient function is a generalized gradient that can be applied to non-smooth but convex objective functions. By determining the subgradient function, effective parameter updates can be performed during the optimization process, making the training process of the uncertain condition prediction training model more stable and efficient, reducing the impact of prediction errors on the optimization results, and improving the accuracy and robustness of scheduling decisions.

[0024] In a specific example, it is necessary to optimize the load prediction model of the data center energy system. First, under the dual optimization training model, the decision loss function of the uncertain condition prediction training model can be calculated, and this function reflects the error between the prediction result and the actual optimal solution. Then, the subgradient function of the decision loss function is determined to use the subgradient for parameter update during the training process. Through subgradient update, the model can better handle non-smooth functions and reduce the impact of prediction errors.

[0025] Step 104: Iteratively update the parameters of the uncertain condition prediction training model through the subgradient function to obtain a trained end-to-end prediction optimization model.

[0026] Among them, the end-to-end prediction optimization model is used to determine the scheduling prediction dual solution of the dual optimization model of the first operation scheduling optimization model according to the boundary condition information corresponding to the first operation scheduling optimization model.

[0027] In some embodiments of the present application, in order to optimize the performance of the uncertain condition prediction training model so that it can accurately predict and optimize the energy system scheduling under uncertain conditions, the parameters of the uncertain condition prediction training model will be iteratively updated through the subgradient function to obtain a trained end-to-end prediction optimization model. The subgradient function can handle the optimization problem of non-smooth functions. Through multiple iterations of update, the prediction error is gradually reduced, and the robustness and decision-making accuracy of the model are improved.

[0028] In a specific example, the subgradient function is used to iteratively update the uncertain condition prediction training model. First, calculate the subgradient of each iteration, and adjust the model parameters according to the direction of the subgradient. Through multiple iterations, the prediction error of the model is gradually reduced, and the prediction result is closer to the actual situation. Finally, a well-trained end-to-end prediction optimization model will be obtained, which can accurately predict the load demand of the data center and provide reliable input for optimizing the scheduling, improving the scheduling decision-making accuracy and robustness of the data center energy system.

[0029] In summary, in the embodiments of the present application, by constructing a dual optimization training model, the uncertain parameters are transferred from the constraint conditions to the objective function, enabling the optimization model to better capture the influence of the uncertain parameters and reducing the dependence on the exact probability distribution. Furthermore, by establishing an uncertain condition prediction training model in the long short-term memory network, the time series characteristics and volatility of the data center load are accurately captured. Based on the boundary condition information of the dual optimization training model, it is ensured that the uncertain condition prediction training model can be consistent with the optimization objective. Then, by determining the subgradient function of the decision loss function, an effective gradient update mechanism is established. Utilizing the characteristic that the subgradient function can handle the optimization problem of non-smooth functions, the training process of the prediction model is made more stable and efficient, reducing the impact of prediction errors on the optimization results and improving the accuracy and robustness of the scheduling decision. Finally, the parameters of the uncertain condition prediction training model are iteratively updated through the subgradient function, realizing the end-to-end integrated training of prediction and optimization, enabling the prediction results to directly serve the optimization objective, enhancing the overall optimization effect of the model in uncertain scenarios, and ensuring the stable operation of the energy system and the efficient allocation of resources. Thus, based on the method of the embodiments of the present application, a closed-loop optimization training process is formed. Through the training and scheduling of the end-to-end prediction optimization model, the efficient scheduling optimization of the data center energy system under uncertain conditions is achieved. Using the dual optimization model and the subgradient update mechanism, the robustness and decision-making accuracy of the model are improved, and the problem of imbalance between the accuracy and optimality of the scheduling strategy during the resource scheduling optimization of the energy system is solved.

[0030] Figure 2 This is another training method for the end-to-end prediction optimization model for the data center energy system provided in this embodiment, which specifically includes the following steps: Step 201, establish a dual optimization training model for the first operation scheduling optimization training model of the data center energy system under the first operation objective.

[0031] The method described in this step has been explained in step 101 and will not be elaborated here.

[0032] Optionally, step 201 includes the following sub-steps: Sub-step 2011, establish a second operation scheduling optimization training model for the data center energy system under the second operation objective, and solve the second operation scheduling optimization training model to obtain the first training solution interval of the data center energy system.

[0033] In some embodiments of the present application, in order to provide a reference solution interval for optimizing the training model of the first operation scheduling, thereby improving the effect of scheduling optimization, a second operation scheduling optimization training model of the data center energy system under the second operation target will be established, and the second operation scheduling optimization training model will be solved to obtain the first training solution interval of the data center energy system. The second operation target is different from the first operation target, providing diverse scheduling references. By solving the second operation scheduling optimization training model, the obtained first training solution interval can be used as the boundary condition of the subsequent optimization model, providing an effective reference basis for subsequent optimization.

[0034] In a specific example, it is desired to optimize the energy scheduling of the data center to maximize the utilization rate of renewable energy. First, a second operation scheduling optimization training model with the goal of maximizing the seasonal target of renewable energy is established. Then, the model is solved to obtain a series of possible scheduling strategies, forming the day-ahead scheduling target interval of the data center energy system. Through this process, some reference solutions are obtained, which will provide valuable boundary conditions in the subsequent optimization process.

[0035] In a specific embodiment of the present application, according to the characteristics of the data center energy system, the coupling and energy conversion between the supply, storage, and consumption of various energies such as hydrogen, electricity, heat, and cold can be described from the perspective of energy supply-demand balance, and a seasonal-day-ahead operation scheduling model of the data center energy system can be established.

[0036] The main body of the data center energy system consists of a photovoltaic panel (SP), a solar collector (ST), a fuel cell (FC), an electrolyzer (EL), a heat pump (HP), an absorption chiller (AC), a refrigeration distribution unit (CD), a hydrogen storage tank (HS), a hot water tank (HW), and a cold water tank (CW), etc. According to the structure of the data center energy system, the equipment operation constraints of each energy supply device and energy storage device in the data center energy system, as well as the energy balance constraints of energies such as hydrogen, electricity, heat, and cold, are set.

[0037] The energy balance constraints include the supply-demand balance constraints of energies such as hydrogen, electricity, heat, and cold:

[0038] In the formula, represents the hydrogen purchase volume of the data center energy system in the th time period; represents the hydrogen production amount of the electrolyzer; represents the hydrogen amount stored / supplied by the hydrogen storage tank; represents the hydrogen consumption amount of the fuel cell; represents the electricity purchase volume in the th time period; and respectively represent the photovoltaic power generation amount and the fuel cell power generation amount; , , and represent the power consumption of the electrolyzer, heat pump, refrigeration distribution unit, and absorption chiller, respectively; represents the electrical load; represents the fuel cell heat recovery efficiency; , and represent the heat output of the fuel cell, solar collector, and heat pump, respectively; represents the heat stored / supplied by the heat storage tank; represents the heat consumption of the absorption chiller; and represent the cooling capacity of the heat pump and absorption chiller, respectively; represents the cold stored / supplied by the cold storage tank; represents the cooling load. If , and are negative, it means energy is stored in the energy storage device; otherwise, it means the energy storage device supplies energy outward.

[0039] In the energy system of the data center, the energy supply devices include photovoltaic panels, solar collectors, fuel cells, electrolyzers, absorption chillers, refrigeration distribution units, etc. Photovoltaic panels mainly generate electricity using solar energy; solar collectors mainly generate heat using solar energy; fuel cells mainly generate electricity using hydrogen energy and generate heat simultaneously during operation; electrolyzers mainly convert electrical energy into hydrogen energy; absorption chillers mainly use the difference between electrical energy and heat energy for refrigeration; the refrigeration distribution unit is used to supply cooling for heat dissipation to each device in the data center. Each device needs to meet its own operation constraints and capacity upper limit constraints.

[0040] The operation constraint expression of the energy supply device is:

[0041] In the formula, and represent the efficiency of converting solar energy into electrical energy and the efficiency of converting solar energy into heat energy, respectively; and represent the construction areas of the photovoltaic panel and the solar collector, respectively; represents the solar radiation amount in the time period; and represent the electrolyzer hydrogen production efficiency by electrolysis and the fuel cell hydrogen power generation efficiency, respectively; and represent the heat production efficiency and the cooling production efficiency of the heat pump, respectively; and respectively represent that the heat pump operates in the heat production mode and the cooling production mode; represents the electric - heat ratio of the fuel cell; represents the heat - to - cold conversion efficiency of the absorption chiller; and respectively represent the power consumption efficiencies of the refrigeration distribution unit and the absorption chiller.

[0042] The upper - limit constraint expression of the energy - supply equipment capacity is:

[0043] In the formula, , , , , , respectively represent the planned capacities of each energy - supply equipment.

[0044] In the data - center energy system, energy - storage equipment includes hydrogen storage tanks, hot - water tanks, cold - water tanks, etc. By using energy - storage equipment to store and release energy across time periods, the data - center energy system can better match energy supply and demand.

[0045] The equipment - operation constraint expression of the energy - storage equipment is:

[0046] In the formula, , and respectively represent the hydrogen storage amount, heat storage amount, and cold storage amount in the time period; , and respectively represent the hydrogen, heat, and cold loss factors.

[0047] The upper - limit constraint expression of the energy - storage equipment capacity is: ,

[0048] In the formula, , , respectively represent the planned capacities of the hydrogen storage tank, hot - water tank, and cold - water tank.

[0049] To effectively regulate cross - seasonal energy storage, the data - center energy - system seasonal - day - ahead operation scheduling model determines the scheduling strategies of energy - storage and non - linear energy - supply equipment through seasonal operation scheduling, and then determines the scheduling strategies of all equipment according to the day - ahead photovoltaic power generation output and the data - center energy consumption load prediction results. The seasonal scheduling model takes the minimum annual operation cost as the objective function.

[0050] The objective - function expression of the seasonal scheduling model is:

[0051] In the formula, represents the scheduling cycle length of the seasonal scheduling model; represents the hydrogen price in the hydrogen market; represents the price of purchasing electricity from the power grid.

[0052] By solving the seasonal scheduling model determined by all the previous constraints, the operating modes of non-linear devices such as heat pumps are obtained , thus obtaining a day-ahead operation scheduling model in the form of linear programming.

[0053] The previous scheduling model takes the minimization of the day-ahead operating cost as the objective function.

[0054] The expression of the objective function is:

[0055] In the formula, represents the day-ahead scheduling cycle length.

[0056] Sub-step 2012: Taking the first training solution interval as the boundary condition and the first operating objective as the optimization objective, a first operating scheduling optimization training model is established.

[0057] Among them, the first operating scheduling optimization training model is constrained according to the uncertain parameters of the data center energy system.

[0058] In some embodiments of the present application, in order to introduce the boundary conditions and optimization objectives in actual operation into the optimization scheduling model to ensure the feasibility and effectiveness of the model, the first training solution interval is taken as the boundary condition and the first operating objective is taken as the optimization objective to establish a first operating scheduling optimization training model. The first operating scheduling optimization training model is constrained according to the uncertain parameters of the data center energy system to ensure that the optimization process can capture the uncertainties and fluctuations in actual operation. This step makes the optimization scheduling model more robust and accurate by introducing the boundary conditions and uncertain parameters in actual operation.

[0059] In a specific example, it is desired to optimize the energy scheduling of the data center to maximize the energy utilization efficiency. First, the optimal solution interval obtained through the seasonal scheduling optimization objective before is taken as the boundary condition, and the day-ahead scheduling objective of the energy utilization efficiency is taken as the first optimization objective to establish a first operating scheduling optimization training model. In the model, the first training solution interval will participate in the subsequent day-ahead scheduling as the uncertain parameter of the data center energy system.

[0060] In the day-ahead operation scheduling model provided by the present application, first, the day-ahead operation scheduling model of the data center energy system needs to be reformulated into the form of an impact matrix:

[0061] In the formula, Represents decision variables related to the balance of electricity and cooling supply and demand; Represents the remaining other decision variables; 、 、 Represents the corresponding coefficient matrix; Represents uncertain parameters such as photovoltaic power generation output, electricity load and cooling load of the data center; Represents the remaining other right-hand vectors.

[0062] Sub-step 2013, perform feasible dual reconstruction on the first operation scheduling optimization training model, and determine the optimization training model obtained according to the feasible dual reconstruction as the dual optimization training model.

[0063] In some embodiments of the present application, in order to solve the uncertain parameters in the optimization model, feasible dual reconstruction will be performed on the first operation scheduling optimization training model, and the optimization training model obtained according to the feasible dual reconstruction will be determined as the dual optimization training model. Feasible dual reconstruction is to reformulate the optimization problem into a dual form to better capture the influence of uncertain parameters. After performing feasible dual reconstruction, the constraint conditions and boundary conditions of the optimization problem are interchanged. By reconstructing and introducing the dual optimization model, the model can be more efficient and stable when dealing with complex constraints and uncertainties.

[0064] In a specific example, it is desired to optimize the energy scheduling of the data center and the uncertainty of renewable energy output and load demand needs to be considered. First, perform feasible dual reconstruction on the first operation scheduling optimization training model containing these uncertain parameters to reformulate the optimization problem into a dual form. Then, determine the dual optimization training model according to the uncertain parameters obtained during the reconstruction process. Through this process, the influence of uncertain parameters can be better captured, and the accuracy and stability of the optimization model in actual operation can be improved.

[0065] In the day-ahead operation scheduling model provided by the present application, perform feasible dual reconstruction on the day-ahead operation scheduling model in matrix form, and establish a dual model of the uncertain parameters in the objective function: , where, 、 Represents the dual variables, and the optimal decision vector expression of the dual model is: , where, S Represents the feasible region; Represents the cost vector, which is ; Represents the decision vector, which is .

[0066] Step 202: Establish an uncertain condition prediction training model for the data center energy system in the long short-term memory network according to the boundary condition information corresponding to the first operation scheduling optimization training model.

[0067] The method shown in this step has been described in step 102 and will not be elaborated here.

[0068] Optionally, step 202 includes the following sub-steps: Sub-step 2021: Initialize a long short-term memory network such that the long short-term memory network includes a first fully connected layer, multiple unidirectional long short-term memory layers, and a second fully connected layer connected in sequence.

[0069] Among them, the second fully connected layer includes a dropout layer and an activation function connected in sequence; the first fully connected layer is used to input the boundary condition sequence of the first operation scheduling optimization training model; the boundary condition sequence is obtained from the boundary condition information corresponding to the first operation scheduling optimization training model according to a preset sequence length; the activation function is used to output a non-negative prediction result of the first operation scheduling optimization training model under the first operation target.

[0070] As Figure 3 shown, it is an example of a long short-term memory network provided according to an embodiment of the present application.

[0071] In some embodiments of the present application, in order to construct a prediction model capable of processing time series data and improve the accuracy of data center energy system load prediction, a long short-term memory (LSTM) network M is initialized such that the LSTM network includes a first fully connected layer X, multiple unidirectional long short-term memory layers Y, and a second fully connected layer Z connected in sequence. The LSTM network is particularly suitable for processing sequence data with time dependence and captures long-term and short-term dependencies in the constrained conditions (c T 、h T )data through memory and forgetting mechanisms. The second fully connected layer includes a dropout layer and an activation function connected in sequence, and the first fully connected layer is used to input the boundary condition sequence f T of the first operation scheduling optimization training model. The boundary condition sequence is obtained from the boundary condition information corresponding to the first operation scheduling optimization training model according to a preset sequence length. The activation function is used to output the non-negative value in the prediction result of the first operation scheduling optimization training model under the first operation target. This structural design enables the prediction model to process complex time series data and improve the accuracy and robustness of the prediction.

[0072] In a specific example, an LSTM network is used to predict the load demand of a data center for the next day. First, an LSTM network with multiple hierarchical structures is initialized, including a first fully connected layer, two unidirectional long short-term memory layers, and a second fully connected layer connected in sequence. The second fully connected layer contains a dropout layer to prevent overfitting and is connected to a linear rectification function (ReLU) as the activation function to ensure that the output result is non-negative. The historical load data and the boundary condition information corresponding to the first operation scheduling optimization training model are input into the LSTM network to generate a boundary condition sequence. With this configuration, the LSTM network can remember and utilize the patterns in the time series and output an accurate prediction result for the future load demand. Finally, the experimenters obtained an LSTM model that can accurately predict the load demand, providing reliable input for the optimization scheduling of the data center.

[0073] In the day-ahead operation scheduling model provided by the present application, since the input features of the prediction model include the time-sharing power and cooling load data of the data center, weather data, and time information, the time-sharing weather data includes temperature, humidity, wind speed, etc., and the time information includes information such as weekdays and holidays. The input of the prediction model is an input sequence composed of the time-sharing input features of the past 7 days: , where, represents the starting moment of the output of the prediction model; represents the length of the input sequence; represents the i-th input feature in the t-th period; represents the length of the input feature. The main structure of the established prediction model is a two-layer unidirectional LSTM network. Specifically, the two-layer unidirectional LSTM network is composed of a fully connected layer, two unidirectional LSTM layers, a Dropout layer, and a fully connected layer with the ReLU function as the activation function in sequence. The first fully connected layer is used to perform feature processing on the input sequence ; the LSTM layer is composed of LSTM cells and is used to extract the sequence information in the input; the Dropout layer is used to suppress model overfitting; the last fully connected layer with the ReLU function as the activation function is used to output the non-negative day-ahead power and cooling load prediction results of the data center.

[0074] Step 203, establish the decision loss function of the uncertain condition prediction training model under the dual optimization training model.

[0075] Among them, the decision loss function is used to characterize the error between the optimal solution of the dual optimization training model obtained according to the uncertain condition prediction training model and the preset optimal training reference solution.

[0076] In some embodiments of the present application, in order to measure the impact of the prediction result on the optimization objective and ensure the optimization effect of the uncertain condition prediction training model, a decision loss function of the uncertain condition prediction training model under the dual optimization training model will be established. The decision loss function is used to characterize the error between the optimal solution of the dual optimization training model obtained according to the uncertain condition prediction training model and the preset optimal training reference solution. By using the decision loss function, the impact of the prediction error on the optimization result can be accurately measured, and a basis for subsequent subgradient calculation can be provided, improving the stability and optimization effect of model training.

[0077] In a specific example, first, under the dual optimization training model, a decision loss function is established. This loss function is used to characterize the error between the predicted load and the actual load, that is, according to the prediction result of the uncertain condition prediction training model, calculate the optimal solution of the dual optimization training model, and compare it with the preset optimal training reference solution to obtain the decision loss value. Through this decision loss value, the performance of the load prediction model can be accurately evaluated, and a reference for further optimization of the model can be provided.

[0078] In the day-ahead operation scheduling model provided by the present application, the expression of the decision loss function is: , where, represents the cost vector predicted value.

[0079] Step 204, according to the convex upper bound of the decision loss function, extend the domain of definition of the decision loss function to establish a surrogate loss function of the decision loss function.

[0080] In some embodiments of the present application, in order to make the decision loss function applicable to a wider range of parameters and improve the training effect and stability of the model, the domain of definition of the decision loss function will be extended according to the convex upper bound of the decision loss function to establish a surrogate loss function of the decision loss function. The convex upper bound is used to describe the boundary conditions of the uncertain condition prediction training model in the worst case. Through the domain extension, the decision loss function can be extended to a larger range, making it more applicable to different scenarios. This enables the surrogate loss function to better reflect the actual situation and improve the robustness and optimization effect of the model.

[0081] In a specific example, it is necessary to establish a more robust decision loss function for the load prediction model of the data center energy system. First, according to the convex upper bound of the decision loss function. Then, based on this convex upper bound, extend the domain of definition of the original decision loss function to establish a surrogate loss function. This surrogate loss function can reflect the error of load prediction within a wider range of parameters, thus showing better robustness and optimization effect in practical applications.

[0082] In the day-ahead operation scheduling model provided in this application, a convex upper bound of the decision loss represented by the input sequence is taken as the surrogate function of the decision loss function: , where represents a convex upper bound of the expression decision loss function, represents the extension radius correction magnification during extension.

[0083] Optionally, step 204 includes the following sub-steps: Sub-step 2041, obtain the load boundary value and the output boundary value of the data center energy system from the boundary condition information corresponding to the first operation scheduling optimization training model in the uncertain condition prediction training model.

[0084] In some embodiments of this application, in order to obtain the boundary condition information required during model training and ensure that the optimization model can accurately reflect the actual situation, the load boundary value and the output boundary value of the data center energy system will be obtained from the boundary condition information corresponding to the first operation scheduling optimization training model in the uncertain condition prediction training model. The load boundary value represents the power load range of the data center in different time periods, and the output boundary value represents the output capacity range of renewable energy. By obtaining these boundary values, accurate data support can be provided for the subsequent extension of the decision loss function domain, enabling the optimization model to perform well under different load and output conditions.

[0085] In a specific example, first obtain the boundary condition information of the uncertain condition prediction training model from historical data, which includes the power load data and renewable energy output data of the data center in the past period. Then, the experimenter extracts the load boundary value and the output boundary value from these data. The load boundary value represents the discrimination value of the power load, and the output boundary value represents the discrimination value of the renewable energy output. By obtaining the boundary values, accurate boundary condition information can be provided for the subsequent optimization model, thereby ensuring that the optimization model can perform well in actual operation and provide a stable and efficient energy scheduling strategy.

[0086] In the day-ahead operation scheduling model provided in this application, the surrogate function has a subgradient, which can be used in the backpropagation process of the data center load prediction model training. The expression of the subgradient is: , where represents the subgradient.

[0087] Sub-step 2042, determine the extension radius correction magnification corresponding to the magnitude relationship according to the magnitude relationship between the load boundary value and the preset load boundary threshold, and between the output boundary value and the preset output boundary threshold.

[0088] In some embodiments of the present application, in order to ensure that the size of the extension radius can accurately reflect the actual situation of load and output during the domain extension process of the decision loss function, the extension radius correction factor corresponding to the size relationship will be determined according to the size relationship between the load boundary value and the preset load boundary threshold, and the output boundary value and the preset output boundary threshold. The load boundary value and the output boundary value reflect the power load and renewable energy output capacity of the data center at different time periods, while the preset boundary threshold is the reference value of the system under different operating conditions. By determining the correction factor, the extension radius can be adjusted under different load and output conditions, making the extension process more accurate and reasonable.

[0089] In a specific example, it is desired to optimize the energy scheduling of the data center to cope with the fluctuations of load demand and renewable energy output. First, the load boundary value and the output boundary value are obtained from historical data and compared with the preset load boundary threshold and output boundary threshold. For example, if the load boundary value exceeds the preset load boundary threshold, it means that the current system load is high, and the experimenter needs to increase the extension radius correction factor. Similarly, if the output boundary value is higher than the preset output boundary threshold, it means that the current renewable energy output is high, and the extension radius correction factor also needs to be increased. Through this adjustment, the appropriate extension radius correction factor can be determined, so as to reasonably extend the domain of the decision loss function in the subsequent steps and ensure that the optimization model can perform well under different load and output conditions.

[0090] Optionally, sub-step 2042 includes the following sub-steps: Sub-step 20421, when the load boundary value is greater than the load boundary threshold and the output boundary value is greater than the output boundary threshold, determine the extension radius correction factor as the first factor.

[0091] In some embodiments of the present application, in order to appropriately correct the extension radius when both the load and the output are high, improve the adaptability and stability of the optimization model, when the load boundary value is greater than the load boundary threshold and the output boundary value is greater than the output boundary threshold, the extension radius correction factor will be determined as the first factor. This process determines the most suitable extension radius correction factor based on the size relationship between the load and output boundary values and thresholds in the actual operation data to ensure the effectiveness and stability of the optimization model under high load and high output conditions. Through this correction, the optimization model can better handle the operating conditions of high load and high output, enhancing the robustness of the system.

[0092] In a specific example, it is necessary to optimize the energy scheduling of the data center and ensure that the model can operate stably under high load and high output conditions. First, obtain the load boundary value and the output boundary value of the data center, and compare them with the preset load boundary threshold and output boundary threshold. It is found that the current load boundary value is greater than the load boundary threshold, and the output boundary value is greater than the output boundary threshold. According to this situation, the extension radius correction magnification is determined as the first magnification. For example, the first magnification can be 5. Through this adjustment, the adaptability and stability of the optimization model under high load and high output conditions are improved, ensuring the effective operation of the energy system of the data center under these conditions.

[0093] In the day-ahead operation scheduling model provided by this application, for the high-load and low-output scenario, at this time, the energy supply is tense and the cost of wrong decisions is high. A larger value is selected to make the upper bound of the loss function tighter and improve the accuracy of the scheduling decision.

[0094] Sub-step 20422, when the load boundary value is less than the load boundary threshold and the output boundary value is less than the output boundary threshold, determine the extension radius correction magnification as the second magnification.

[0095] In some embodiments of this application, in order to appropriately correct the extension radius when both the load and the output are low to ensure the robustness and adaptability of the optimization model, when the load boundary value is less than the load boundary threshold and the output boundary value is less than the output boundary threshold, the extension radius correction magnification will be determined as the second magnification. This process determines the most suitable extension radius correction magnification based on the magnitude relationship between the load and output boundary values and thresholds in the actual operation data to ensure the effectiveness and stability of the optimization model under low load and low output conditions. Through this correction, the optimization model can better handle the operating conditions of low load and low output, enhancing the robustness of the system.

[0096] In a specific example, it is necessary to optimize the energy scheduling of the data center and ensure that the model can operate stably under low load and low output conditions. First, obtain the load boundary value and the output boundary value of the data center, and compare them with the preset load boundary threshold and output boundary threshold. It is found that the current load boundary value is less than the load boundary threshold, and the output boundary value is less than the output boundary threshold. According to this situation, the extension radius correction magnification is determined as the second magnification. For example, the second magnification can be 2. Through this adjustment, the adaptability and stability of the optimization model under low load and low output conditions are improved, ensuring the effective operation of the energy system of the data center under these conditions.

[0097] In the day-ahead operation scheduling model provided by this application, for the low-load and high-output scenario, where the energy supply is sufficient at this time and the system has high flexibility, a smaller value is selected to ensure the convexity of the loss function and the numerical stability of the subgradient.

[0098] Sub-step 20423: In cases other than those corresponding to sub-step 20421 or sub-step 20422, determine the extension radius correction factor as the third factor.

[0099] Among them, the first factor is greater than the third factor; the third factor is greater than the second factor; the second factor is greater than 1.

[0100] In some embodiments of this application, in order to appropriately correct the extension radius when the load and output do not simultaneously meet high or low conditions, and to ensure the stability and adaptability of the optimization model, in cases other than those corresponding to sub-step 20421 or sub-step 20422, determine the extension radius correction factor as the third factor. The third factor is used to handle the case where the load and output are at a medium level. Through an appropriate correction factor, the optimization model can still perform well under non-extreme load and output conditions. Through this correction, the optimization model can better adapt to various situations in actual operation, improving the robustness and stability of the system.

[0101] In a specific example, it is necessary to optimize the energy scheduling of a data center and ensure that the model can operate stably when the load and output are at a medium level. First, obtain the load boundary value and output boundary value of the data center, and compare them with the preset load boundary threshold and output boundary threshold. It is found that the current load boundary value and output boundary value do not meet the high-load and high-output or low-load and low-output conditions. According to this situation, determine the extension radius correction factor as the third factor. For example, the third factor can be 3. Through this adjustment, the adaptability and stability of the optimization model under medium load and output conditions are improved, ensuring the effective operation of the energy system of the data center under these conditions.

[0102] In the day-ahead operation scheduling model provided by this application, for the high-load and high-output scenario and the low-load and low-output scenario, where the system is in a relatively balanced state at this time and the cost of incorrect decisions is moderate, select the intermediate value to balance the upper bound tightness and the numerical stability of the subgradient.

[0103] Sub-step 2043: Extend the domain of definition of the decision loss function according to the extension radius correction factor to establish an alternative loss function of the decision loss function.

[0104] In some embodiments of the present application, in order to make the decision loss function more widely applicable to different operating conditions, thereby improving the robustness and effectiveness of the optimization model, the domain of the decision loss function will be extended according to the extension radius correction factor to establish an alternative loss function for the decision loss function. The extension radius correction factor is determined based on the magnitude relationship between the load boundary value and the preset load boundary threshold, as well as the output boundary value and the preset output boundary threshold, and is used to adjust the range of the decision loss function during the domain extension process. By performing domain extension, the alternative loss function can better adapt to the changes in actual operation, improving the applicability and optimization effect of the model under different load and output conditions.

[0105] In a specific example, it is desired to optimize the energy scheduling of a data center and ensure that the model can operate effectively under various load and output conditions. First, according to the extension radius correction factor, the extension range of the decision loss function is calculated. Then, the experimenter extends the domain of the decision loss function to establish a new alternative loss function. For example, if the correction factor is 2, the experimenter will expand the extension range with a radius of 2 based on the original decision loss function. Through this process, the new alternative loss function can more flexibly handle different operating conditions and provide more reliable optimization results. Finally, the experimenter obtains an alternative loss function with good optimization effect, improving the scheduling decision accuracy and robustness of the data center energy system.

[0106] In the day-ahead operation scheduling model provided by the present application, the value determined according to the scenario will be used to correct the alternative function:

[0107] Step 205, determine the subgradient function of the established alternative loss function as the subgradient function of the decision loss function.

[0108] In some embodiments of the present application, in order to use the gradient function of the alternative loss function to handle non-smooth optimization problems, thereby improving the optimization effect of the uncertain condition prediction training model, the subgradient function of the established alternative loss function will be determined as the subgradient function of the decision loss function. The subgradient is a generalized form of the gradient and is applicable to non-smooth but convex objective functions. By using the gradient function of the alternative loss function as the subgradient function of the decision loss function, effective parameter updates can be performed during the optimization process, improving the training stability and optimization effect of the uncertain condition prediction training model.

[0109] In a specific example, it is necessary to optimize the load prediction model of the data center energy system. First, according to the uncertain conditions, the surrogate loss function established by predicting the training model and the dual optimization training model is calculated, and its sub-gradient function is calculated. Then, this gradient function is determined as the sub-gradient function of the decision loss function. By using the sub-gradient to update the parameters during the training process, the model can handle non-smooth optimization problems and reduce the impact of prediction errors.

[0110] In the day-ahead operation scheduling model provided in this application, the value determined according to the scenario is used to correct the surrogate loss function: .

[0111] Step 206, iteratively update the parameters of the uncertain condition prediction training model through the sub-gradient function to obtain a trained end-to-end prediction optimization model.

[0112] Among them, the end-to-end prediction optimization model is used to determine the scheduling prediction dual solution of the dual optimization model of the first operation scheduling optimization model according to the boundary condition information corresponding to the first operation scheduling optimization model.

[0113] The method shown in this step has been described in step 104 and will not be elaborated here.

[0114] In summary, in the embodiments of the present application, by constructing a dual optimization training model, the uncertain parameters are transferred from the constraint conditions to the objective function, enabling the optimization model to better capture the influence of the uncertain parameters and reducing the dependence on the exact probability distribution. Furthermore, by establishing an uncertain condition prediction training model in the long short-term memory network, the time series characteristics and volatility of the data center load are accurately captured. Based on the boundary condition information of the dual optimization training model, it is ensured that the uncertain condition prediction training model is consistent with the optimization objective. Then, by determining the subgradient function of the decision loss function, an effective gradient update mechanism is established. Utilizing the characteristic that the subgradient function can handle the optimization problem of non-smooth functions, the training process of the prediction model is made more stable and efficient, reducing the impact of prediction errors on the optimization results and improving the accuracy and robustness of the scheduling decision. Finally, by iteratively updating the parameters of the uncertain condition prediction training model using the subgradient function, an end-to-end integrated training of prediction and optimization is achieved, enabling the prediction results to directly serve the optimization objective, enhancing the overall optimization effect of the model in uncertain scenarios, and ensuring the stable operation of the energy system and the efficient allocation of resources. Thus, based on the method of the embodiments of the present application, a closed-loop optimization training process is formed. Through the training and scheduling of the end-to-end prediction optimization model, the efficient scheduling optimization of the data center energy system under uncertain conditions is realized. Using the dual optimization model and the subgradient update mechanism, the robustness and decision-making accuracy of the model are improved, and the problem of imbalance between the accuracy and optimality of the scheduling strategy during the resource scheduling optimization of the energy system is solved.

[0115] Based on the end-to-end prediction optimization model trained according to the embodiments of the present application, as Figure 4 shown, the embodiments of the present application also provide an end-to-end prediction optimization scheduling method for a data center energy system, which specifically includes the following steps: Step 301, establish a dual optimization model of the first operation scheduling optimization model of the data center energy system under the first operation objective.

[0116] In some embodiments of the present application, in order to transform the operation scheduling problem of the data center energy system to make it easier to solve and optimize, a dual optimization model of the first operation scheduling optimization model of the data center energy system under the first operation objective will be established. The dual optimization model is a method of reformulating the optimization problem into a dual form, which can simplify the calculation and better handle uncertain parameters. By establishing the dual optimization model, a complex optimization problem can be transformed into a more tractable form, improving the optimization efficiency and solution stability.

[0117] In a specific example, it is desired to optimize the energy system scheduling of a data center to meet the day-ahead scheduling. First, determine the day-ahead scheduling objectives of the data center. Then, establish a first operation scheduling optimization model that includes these objectives and constraints. Next, perform a dual transformation on this optimization model to form a dual optimization model. Through this process, the optimization problem can be simplified, the solution efficiency can be improved, and the impact of uncertain parameters can be better captured.

[0118] Step 302: Input the boundary condition information corresponding to the first operation scheduling optimization training model into the end-to-end prediction optimization model trained by the training method of the end-to-end prediction optimization model for the data center energy system disclosed in the embodiments of the present application, so as to obtain the scheduling prediction dual solution of the dual optimization model.

[0119] In some embodiments of the present application, in order to use the trained end-to-end prediction optimization model to predict the scheduling decision in actual operation, the boundary condition information corresponding to the first operation scheduling optimization training model will be input into the end-to-end prediction optimization model trained by the training method of the end-to-end prediction optimization model for the data center energy system disclosed in the embodiments of the present application, so as to obtain the scheduling prediction dual solution of the dual optimization model. The boundary condition information includes various limitations and conditions in the actual operation of the data center energy system. By inputting this information into the trained end-to-end prediction optimization model, the model can generate the scheduling prediction dual solution of the dual optimization model, thereby guiding the actual scheduling decision. This step ensures that the prediction model can provide accurate scheduling predictions according to the actual operation situation, improving the practicability and reliability of the optimization model.

[0120] In a specific example, it is desired to use the trained end-to-end prediction optimization model to predict the power scheduling strategy of the data center for the next day. First, obtain the boundary condition information corresponding to the first operation scheduling optimization training model, including historical load data, weather forecasts, and equipment operation constraints, etc. Then, input these boundary condition information into the trained end-to-end prediction optimization model, and the model generates the scheduling prediction dual solution of the dual optimization model according to the input information. Through this process, the scheduling prediction results of the data center energy system will be obtained, and these results can be used to guide the resource scheduling of the data center for the next day to ensure the efficient operation of the system and the optimized utilization of energy.

[0121] Step 303: Use the scheduling prediction dual solution as the boundary condition to solve the first operation scheduling optimization model to obtain the scheduling prediction solution of the first operation scheduling optimization model, and use the scheduling prediction solution as the scheduling configuration to perform resource scheduling on the data center energy system.

[0122] In some embodiments of the present application, in order to apply the output of the prediction model to actual scheduling optimization and achieve effective scheduling of the data center energy system, the scheduling prediction dual solution is used as a boundary condition to solve the first operation scheduling optimization model to obtain the scheduling prediction solution of the first operation scheduling optimization model. Then, resource scheduling of the data center energy system is performed according to the scheduling prediction solution. The scheduling prediction dual solution is generated based on the trained end-to-end prediction optimization model and represents the prediction result of the actual scheduling problem. By using these prediction results as boundary conditions, the optimization model can better reflect the actual situation, thereby providing an accurate scheduling plan. Finally, the efficient operation of the data center energy system and the optimal allocation of resources are ensured.

[0123] In a specific example, it is desired to utilize the prediction results of the end-to-end prediction optimization model, and the scheduling prediction solutions have been obtained through the previous steps. These prediction solutions contain the scheduling information of the data center in future time periods. Next, based on these scheduling prediction solutions, a specific resource scheduling plan can be formulated and implemented to ensure the stable operation of the data center. Through this process, the results of the prediction model can be effectively utilized, improving the operation efficiency and resource utilization rate of the data center energy system.

[0124] In summary, in the embodiments of the present application, by inputting the boundary condition information into the trained end-to-end prediction optimization model, the scheduling prediction dual solution is generated in real time, ensuring that the optimization model can dynamically adapt to the operation status and load demand of the data center energy system, improving the flexibility and response speed of the system; and then, by using the scheduling prediction dual solution as a boundary condition to solve the first operation scheduling optimization model, the scheduling prediction solution is finally obtained, realizing the optimal scheduling of the data center energy system, enabling the system to achieve the optimal resource allocation and balance under uncertain conditions, and ensuring the stable operation and efficient utilization of the energy system. Thus, based on the method of the embodiments of the present application, a closed-loop optimization training process is formed. Through the training and scheduling of the end-to-end prediction optimization model, the efficient scheduling optimization of the data center energy system under uncertain conditions is realized. By using the dual optimization model and the subgradient update mechanism, the robustness and decision-making accuracy of the model are improved, and the problem of imbalance between the accuracy and optimality of the scheduling strategy during the resource scheduling optimization of the energy system is solved.

[0125] As Figure 5 shown, it is the complete process of prediction optimization scheduling in the embodiments of the present application, specifically including the following processes: S1. Establish an operation scheduling model for the data center energy system: According to the characteristics of the energy system of the data center, an operation scheduling model is constructed to better manage and optimize energy use. To provide a basic scheduling model for subsequent optimization and prediction.

[0126] S2. Perform a feasible dual reconstruction on the operation scheduling model to obtain the operation scheduling dual model: Reconstruct the initial scheduling model for feasibility and duality to ensure the accuracy and feasibility of the model. This can optimize the model structure, enabling it to better handle uncertain parameters, and improving the solution efficiency and stability.

[0127] S3. Establish an LSTM-based prediction model for the uncertain quantities in the data center: Based on the hourly photovoltaic power generation output, data center electricity and cooling load data, weather data, and time information, establish an uncertainty prediction model based on the long short-term memory (LSTM) network to capture the dependencies in time series data using the LSTM network and provide more accurate load and output predictions.

[0128] S4. Establish a decision loss function for training the data center uncertain quantity prediction model: Based on the dual model, establish a decision loss function for training the uncertainty prediction model to ultimately ensure that the prediction model can be trained according to the optimization objective and improve the prediction accuracy of the model.

[0129] S5. Obtain the subgradient of the decision loss function and train the prediction model end-to-end: By calculating the subgradient of the decision loss function, perform end-to-end training to optimize the parameters of the uncertainty prediction model. In this way, through subgradient updates, the prediction error is reduced, and the robustness and decision-making accuracy of the model are improved.

[0130] S6. Solve the original operation scheduling model to obtain the operation scheduling strategy: Using the current photovoltaic power generation output and the predicted results of the data center electricity and cooling loads as the parameters of the original operation scheduling model, solve the original operation scheduling model to obtain the operation scheduling strategy for the data center energy system. Finally, based on the prediction results, generate an optimized scheduling strategy to ensure the efficient operation and resource optimization configuration of the data center energy system.

[0131] As Figure 6 shown, it is the data flow process of prediction and optimization scheduling in the embodiment of this application, specifically including the following processes: R1. Optimization model construction: R11. Scheduling model construction: According to the characteristics of the data center energy system, construct an initial scheduling model. This model defines the operation objectives and constraints of the energy system, provides the basis for the optimization model, and offers a reference framework for the subsequent dual model construction and optimization processes; R12. Dual model construction: Perform a dual reconstruction on the initial scheduling model to form a dual optimization model. This reconstruction can effectively handle uncertain parameters and complex constraints, simplify the optimization calculation process, and improve the robustness and solution efficiency of the model.

[0132] R2. End-to-end training: R21, Training data input: Input the training data containing historical data, weather data, and time information. These data are used to capture the time-series characteristics of load and output, provide necessary data support for the prediction model, and improve the accuracy of prediction; R22, Construction of the day-ahead model: Construct a day-ahead prediction model for the uncertainties in the data center based on the training data. This model can predict the load demand and energy output in the future time period. Technical effect: Provide accurate prediction results of load and output, and lay a foundation for subsequent scheduling optimization; R23, Iterative training: Use the end-to-end decision loss function to perform iterative training on the end-to-end prediction optimization model established based on the dual model established in R12. By calculating the subgradient, continuously update the model parameters, reduce the prediction error, enhance the stability and robustness of the model, and improve the decision-making accuracy of prediction.

[0133] R3, Prediction optimization: R31, Data input: Input the real-time data containing historical data, weather data, and time information. These data reflect the current state in actual operation, provide real-time input data for the model, and improve the timeliness of prediction; R32, Construction of the day-ahead model: Update the day-ahead prediction model based on the input real-time data. This model is used to predict the load demand and energy output in the future time period, provide updated prediction results, and ensure the timeliness and accuracy of the scheduling strategy.

[0134] R33, Solving the scheduling model: Based on the prediction results, solve the initial scheduling model. This process generates the actual scheduling prediction solution, provides specific scheduling strategies, and ensures the efficient operation of the energy system and the optimal allocation of resources; R34, Execution of the scheduling strategy: According to the generated scheduling prediction solution, perform resource scheduling on the data center energy system. This process realizes the implementation of the scheduling optimization strategy, ensures the stable operation of the data center energy system, and improves the energy utilization efficiency and operation benefits of the system.

[0135] As Figure 7 shown, the embodiment of the present application also discloses a training device 40 for an end-to-end prediction optimization model for a data center energy system, including: The training dual construction module 401 is used to establish a dual optimization training model for the first operation scheduling optimization training model of the data center energy system under the first operation target; The training prediction construction module 402 is used to establish an uncertain condition prediction training model for the data center energy system in the long short-term memory network according to the boundary condition information corresponding to the first operation scheduling optimization training model; The subgradient construction module 403 is used to determine the subgradient function of the decision loss function of the uncertain condition prediction training model under the dual optimization training model; The model training module 404 is used to iteratively update the parameters of the uncertain condition prediction training model through the subgradient function to obtain a trained end-to-end prediction optimization model; the end-to-end prediction optimization model is used to determine the scheduling prediction dual solution of the dual optimization model of the first operation scheduling optimization model according to the boundary condition information corresponding to the first operation scheduling optimization model.

[0136] Optionally, the training dual construction module 401 includes: The first training solution solving sub-module is used to establish a second operation scheduling optimization training model of the data center energy system under the second operation target, and solve the second operation scheduling optimization training model to obtain the first training solution interval of the data center energy system; The first training model establishing sub-module is used to establish a first operation scheduling optimization training model with the first training solution interval as the boundary condition and the first operation target as the optimization target; the first operation scheduling optimization training model is constrained according to the uncertain parameters of the data center energy system; The dual reconstruction sub-module is used to perform a feasible dual reconstruction on the first operation scheduling optimization training model, and determine the optimized training model obtained according to the feasible dual reconstruction as the dual optimization training model.

[0137] Optionally, the subgradient construction module 403 includes: The decision loss sub-module is used to establish a decision loss function of the uncertain condition prediction training model under the dual optimization training model; the decision loss function is used to characterize the error between the optimal solution of the dual optimization training model obtained according to the uncertain condition prediction training model and the preset optimal training reference solution; The surrogate loss sub-module is used to extend the domain of definition of the decision loss function according to the convex upper bound of the decision loss function to establish a surrogate loss function of the decision loss function; The subgradient sub-module is used to determine the subgradient function of the established surrogate loss function as the subgradient function of the decision loss function.

[0138] Optionally, the surrogate loss sub-module includes: The boundary value extraction unit is used to obtain the load boundary value and the output boundary value of the data center energy system from the boundary condition information corresponding to the uncertain condition prediction training model and the first operation scheduling optimization training model; The magnification selection unit is used to determine the extension radius correction magnification corresponding to the size relationship according to the size relationship between the load boundary value and the preset load boundary threshold, and the output boundary value and the preset output boundary threshold; An extension unit for extending the domain of definition of the decision loss function according to the extension radius correction magnification to establish an alternative loss function of the decision loss function.

[0139] Optionally, the magnification selection unit includes: The first magnification selection sub-unit is used to determine the extension radius correction magnification as the first magnification when the load boundary value is greater than the load boundary threshold and the output boundary value is greater than the output boundary threshold. The second magnification selection sub-unit is used to determine the extension radius correction magnification as the second magnification when the load boundary value is less than the load boundary threshold and the output boundary value is less than the output boundary threshold. The third magnification selection sub-unit is used to determine the extension radius correction magnification as the third magnification in cases other than when the load boundary value is greater than the load boundary threshold and the output boundary value is greater than the output boundary threshold, or when the load boundary value is less than the load boundary threshold and the output boundary value is less than the output boundary threshold; the first magnification is greater than the third magnification; the third magnification is greater than the second magnification; the second magnification is greater than 1.

[0140] Optionally, the training prediction construction module 402 includes: The long short-term memory network construction sub-module is used to initialize a long short-term memory network so that the long short-term memory network includes a first fully connected layer, multiple unidirectional long short-term memory layers, and a second fully connected layer connected in sequence; the second fully connected layer includes a dropout layer and an activation function connected in sequence; the first fully connected layer is used to input the boundary condition sequence of the first operation scheduling optimization training model; the boundary condition sequence is obtained from the boundary condition information corresponding to the first operation scheduling optimization training model according to a preset sequence length; the activation function is used to output a non-negative prediction result of the first operation scheduling optimization training model under the first operation target.

[0141] As Figure 8 shown, an end-to-end prediction optimization scheduling device 50 for a data center energy system disclosed in an embodiment of the present application includes: The dual construction module 501 is used to establish a dual optimization model of the first operation scheduling optimization model of the data center energy system under the first operation target. The parameter input module 502 is used to input the boundary condition information corresponding to the first operation scheduling optimization training model into the end-to-end prediction optimization model trained according to the training method of the end-to-end prediction optimization model for a data center energy system disclosed in an embodiment of the present application to obtain a scheduling prediction dual solution for the dual optimization model. The solution scheduling module 503 is configured to use the scheduling prediction dual solution as a boundary condition to solve the first operating scheduling optimization model, so as to obtain a scheduling prediction solution of the first operating scheduling optimization model, and use the scheduling prediction solution as a scheduling configuration to perform resource scheduling on the data center energy system.

[0142] In summary, in the embodiments of the present application, by constructing a dual optimization training model, the uncertain parameters are transferred from the constraint conditions to the objective function, enabling the optimization model to better capture the influence of the uncertain parameters and reducing the dependence on the exact probability distribution. Furthermore, by establishing an uncertain condition prediction training model in the long short-term memory network, the time series characteristics and volatility of the data center load are accurately captured. Based on the boundary condition information of the dual optimization training model, it is ensured that the uncertain condition prediction training model is consistent with the optimization objective. Then, by determining the subgradient function of the decision loss function, an effective gradient update mechanism is established. Utilizing the property that the subgradient function can handle the optimization problem of non-smooth functions, the training process of the prediction model is made more stable and efficient, reducing the impact of prediction errors on the optimization results and improving the accuracy and robustness of the scheduling decision. Finally, by iteratively updating the parameters of the uncertain condition prediction training model using the subgradient function, an end-to-end integrated training of prediction and optimization is achieved, enabling the prediction results to directly serve the optimization objective, enhancing the overall optimization effect of the model in uncertain scenarios, and ensuring the stable operation of the energy system and the efficient allocation of resources. Thus, based on the method of the embodiments of the present application, a closed-loop optimization training process is formed. Through the training and scheduling of the end-to-end prediction optimization model, efficient scheduling optimization of the data center energy system under uncertain conditions is realized. By using the dual optimization model and the subgradient update mechanism, the robustness and decision accuracy of the model are enhanced, and the problem of imbalance between the accuracy and optimality of the scheduling strategy during the resource scheduling optimization of the energy system is solved.

[0143] The embodiments of the present application further provide a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is executed by a processor, it implements each process of the above-mentioned training method of the end-to-end prediction optimization model for the data center energy system or the embodiments of the end-to-end prediction optimization scheduling method for the data center energy system, and can achieve the same technical effects. To avoid repetition, it will not be elaborated here. Among them, the computer-readable storage medium is, for example, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disc, etc.

[0144] Figure 9It is a block diagram of an electronic device 700 provided by an embodiment of the present application. For example, the electronic device 700 may be a mobile phone, a computer, a digital broadcast terminal, a messaging device, a game console, a tablet device, a medical device, a fitness device, a personal digital assistant, etc.

[0145] Referring to Figure 9 , the electronic device 700 may include one or more of the following components: a processing component 702, a memory 704, a power component 706, a multimedia component 708, an audio component 710, an input / output (I / O) interface 712, a sensor component 714, and a communication component 716.

[0146] The processing component 702 generally controls the overall operation of the electronic device 700, such as operations associated with display, telephone calls, data communication, camera operations, and recording operations. The processing component 702 may include one or more processors 720 to execute instructions to complete all or part of the steps of the above-mentioned end-to-end prediction optimization model training method for a data center energy system or the end-to-end prediction optimization scheduling method for a data center energy system. In addition, the processing component 702 may include one or more modules to facilitate the interaction between the processing component 702 and other components. For example, the processing component 702 may include a multimedia module to facilitate the interaction between the multimedia component 708 and the processing component 702.

[0147] The memory 704 is used to store various types of data to support the operation of the electronic device 700. Examples of such data include instructions for any application or method operating on the electronic device 700, contact data, phone book data, messages, pictures, multimedia, etc. The memory 704 may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, a magnetic disk, or an optical disk.

[0148] The power component 706 provides power to various components of the electronic device 700. The power component 706 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the electronic device 700.

[0149] The multimedia component 708 includes a screen that provides an output interface between the electronic device 700 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touch screen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operations. In some embodiments, the multimedia component 708 includes a front camera and / or a rear camera. When the electronic device 700 is in an operating mode, such as a shooting mode or a multimedia mode, the front camera and / or the rear camera can receive external multimedia data. Each of the front camera and the rear camera can be a fixed optical lens system or have a focal length and optical zoom capabilities.

[0150] The audio component 710 is used to output and / or input audio signals. For example, the audio component 710 includes a microphone (MIC) that is used to receive external audio signals when the electronic device 700 is in an operating mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signals can be further stored in the memory 704 or transmitted via the communication component 716. In some embodiments, the audio component 710 further includes a speaker for outputting audio signals.

[0151] The I / O interface 712 provides an interface between the processing component 702 and a peripheral interface module, which can be a keyboard, a click wheel, buttons, etc. These buttons can include but are not limited to: a home button, a volume button, a power button, and a lock button.

[0152] The sensor component 714 includes one or more sensors for providing status assessments of various aspects of the electronic device 700. For example, the sensor component 714 can detect the on / off state of the electronic device 700, the relative positioning of components, such as the display and the keypad of the electronic device 700. The sensor component 714 can also detect a change in the position of the electronic device 700 or a component of the electronic device 700, the presence or absence of user contact with the electronic device 700, the orientation or acceleration / deceleration of the electronic device 700, and a change in the temperature of the electronic device 700. The sensor component 714 can include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor component 714 can also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor component 714 can further include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

[0153] The communication component 716 is used to facilitate communication between the electronic device 700 and other devices in a wired or wireless manner. The electronic device 700 can access a communication standard-based wireless network, such as WiFi, a carrier network (such as 2G, 3G, 4G, or 7G), or a combination thereof. In an exemplary embodiment, the communication component 716 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 716 further includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra Wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0154] In an exemplary embodiment, the electronic device 700 can be implemented by one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components, for implementing the training method of the end-to-end prediction optimization model for a data center energy system or the end-to-end prediction optimization scheduling method for a data center energy system provided in the embodiments of the present application.

[0155] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is further provided, such as a memory 704 including instructions, and the above instructions can be executed by the processor 720 of the electronic device 700 to complete the above-mentioned training method of the end-to-end prediction optimization model for a data center energy system or the end-to-end prediction optimization scheduling method for a data center energy system. For example, the non-transitory storage medium can be a ROM, Random Access Memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0156] Figure 10 is a block diagram of an electronic device 800 shown according to an exemplary embodiment. For example, the electronic device 800 can be provided as a server. Referring to Figure 8 , the electronic device 800 includes a processing component 822, which further includes one or more processors, and memory resources represented by a memory 832 for storing instructions executable by the processing component 822, such as application programs. The application programs stored in the memory 832 can include one or more modules each corresponding to a set of instructions. In addition, the processing component 822 is configured to execute instructions to perform the training method of the end-to-end prediction optimization model for a data center energy system or the end-to-end prediction optimization scheduling method for a data center energy system provided in the embodiments of the present application.

[0157] The electronic device 800 may further include a power supply component 826 configured to perform power management of the electronic device 800, a wired or wireless network interface 850 configured to connect the electronic device 800 to a network, and an input / output (I / O) interface 858. The electronic device 800 may operate based on an operating system stored in the memory 832, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSD TM or the like.

[0158] An embodiment of the present application also provides a computer program product, including a computer program, which when executed by a processor, implements a training method for an end-to-end prediction optimization model for a data center energy system or an end-to-end prediction optimization scheduling method for a data center energy system.

[0159] Those skilled in the art will readily conceive of other embodiments of the present application after considering the specification and practicing the application disclosed herein. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include well-known knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and embodiments are only regarded as exemplary, and the true scope and spirit of the present application are pointed out by the following claims.

[0160] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present application is only limited by the appended claims.

[0161] Each embodiment in this specification is described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0162] The training method for an end-to-end prediction optimization model for a data center energy system or the end-to-end prediction optimization scheduling method for a data center energy system provided herein is not inherently related to any specific computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. It is obvious from the above description how to construct the structure required for a system with the solution of the present application. In addition, the present application is not directed to any specific programming language. It should be understood that the content of the present application described herein can be implemented using various programming languages, and the description of the specific language above is for disclosing the best mode of the present application.

[0163] In the description provided herein, numerous specific details are set forth. However, it will be understood that embodiments of the present application may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail so as not to obscure the understanding of this description.

[0164] Those skilled in the art will appreciate that the modules in the devices in the embodiments can be adaptively changed and disposed in one or more devices different from that of the embodiment. The modules or units or components in the embodiments can be combined into one module or unit or component, and in addition, they can be divided into multiple sub-modules or sub-units or sub-components. Except that at least some of such features and / or processes or units are mutually exclusive, any combination can be adopted to combine all the features disclosed in this description (including the accompanying claims, abstract and drawings) and all the processes or units of any method or device so disclosed. Unless otherwise expressly stated, each feature disclosed in this description (including the accompanying claims, abstract and drawings) can be replaced by an alternative feature providing the same, equivalent or similar purpose.

[0165] It should be noted that the above embodiments illustrate rather than limit the present application, and those skilled in the art can design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in the claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present application can be implemented by means of hardware including several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices can be embodied by the same item of hardware. The use of the words first, second, and third, etc. does not denote any order. These words can be interpreted as names.

[0166] The above are only the preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are included within the protection scope of the present invention.

Claims

1. A training method for an end-to-end prediction optimization model for a data center energy system, characterized in that: include: Establishing a dual optimization training model of the first operation scheduling optimization training model of the data center energy system under the first operation target; According to the boundary condition information corresponding to the first operation scheduling optimization training model, establishing an uncertain condition prediction training model of the data center energy system in the long short-term memory network; Determine a subgradient function of a decision loss function of the uncertain condition prediction training model under the dual optimization training model; The parameters of the uncertain condition prediction training model are iteratively updated through the subgradient function to obtain a trained end-to-end prediction optimization model; the end-to-end prediction optimization model is used to determine the scheduling prediction dual solution of the dual optimization model of the first operation scheduling optimization model based on the boundary condition information corresponding to the first operation scheduling optimization model.

2. The training method of the end-to-end prediction optimization model for the data center energy system according to claim 1, characterized in that: The dual optimization training model for establishing a first operation scheduling optimization training model for the data center energy system under the first operation target includes: Establishing a second operation scheduling optimization training model for the data center energy system under a second operation target, and solving the second operation scheduling optimization training model to obtain a first training solution interval for the data center energy system; The first training solution interval is used as a boundary condition and the first operation target is used as an optimization target to establish the first operation scheduling optimization training model; the first operation scheduling optimization training model is constrained according to the uncertain parameters of the data center energy system; A feasible dual reconstruction is performed on the first operation scheduling optimization training model, and the optimization training model obtained according to the feasible dual reconstruction is determined as the dual optimization training model.

3. The training method of the end-to-end prediction optimization model for the data center energy system according to claim 1, characterized in that: The determining of the subgradient function of the decision loss function of the uncertain condition prediction training model under the dual optimization training model includes: Establishing a decision loss function of the uncertain condition prediction training model under the dual optimization training model; the decision loss function is used to characterize the error between the optimal solution of the dual optimization training model obtained according to the uncertain condition prediction training model and the preset optimal training reference solution; According to the convex upper bound of the decision loss function, the domain of the decision loss function is extended to establish a replacement loss function of the decision loss function; The established subgradient function of the replacement loss function is determined as the subgradient function of the decision loss function.

4. The training method of the end-to-end prediction optimization model for the data center energy system according to claim 3 is characterized in that: The step of extending the domain of the decision loss function according to the convex upper bound of the decision loss function to establish a replacement loss function for the decision loss function includes: Obtaining the load boundary value and the output boundary value of the data center energy system from the boundary condition information corresponding to the first operation scheduling optimization training model in the uncertain condition prediction training model; According to the magnitude relationship between the load boundary value and the preset load boundary threshold, and between the output boundary value and the preset output boundary threshold, determining the extension radius correction factor corresponding to the magnitude relationship; The decision loss function is extended in its definition domain according to the extension radius correction factor to establish a replacement loss function for the decision loss function.

5. The training method of the end-to-end prediction optimization model for the data center energy system according to claim 4 is characterized in that: The determining, according to the magnitude relationship between the load boundary value and the preset load boundary threshold, and between the output boundary value and the preset output boundary threshold, a correction factor of the extended radius corresponding to the magnitude relationship comprises: When the load boundary value is greater than the load boundary threshold value, and the output boundary value is greater than the output boundary threshold value, determining the extension radius correction factor as a first factor; When the load boundary value is less than the load boundary threshold value, and the output boundary value is less than the output boundary threshold value, determining the extension radius correction factor as a second factor; When the load boundary value is greater than the load boundary threshold and the output boundary value is greater than the output boundary threshold, or when the load boundary value is less than the load boundary threshold and the output boundary value is less than the output boundary threshold, the extension radius correction factor is determined to be a third factor; the first factor is greater than the third factor; the third factor is greater than the second factor; and the second factor is greater than 1.

6. The training method of the end-to-end prediction optimization model for the data center energy system according to claim 1, characterized in that: The step of establishing the uncertain condition prediction training model of the data center energy system in the long short-term memory network according to the boundary condition information corresponding to the first operation scheduling optimization training model includes: Initialize a long short-term memory network so that the long short-term memory network includes a first fully connected layer, multiple unidirectional long short-term memory layers, and a second fully connected layer that are connected in sequence; the second fully connected layer includes a random inactivation layer and an activation function that are connected in sequence; the first fully connected layer is used to input the boundary condition sequence of the first operation scheduling optimization training model; the boundary condition sequence is obtained from the boundary condition information corresponding to the first operation scheduling optimization training model according to a preset sequence length; the activation function is used to output a non-negative prediction result of the first operation scheduling optimization training model under the first operation target.

7. An end-to-end prediction optimization scheduling method for a data center energy system, characterized in that: include: Establishing a dual optimization model of the first operation scheduling optimization model of the data center energy system under the first operation target; Inputting the boundary condition information corresponding to the first operation scheduling optimization training model into the end-to-end prediction optimization model trained by the training method of the end-to-end prediction optimization model for the data center energy system according to any one of claims 1 to 6 to obtain a scheduling prediction dual solution for the dual optimization model; The scheduling prediction dual solution is used as a boundary condition to solve the first operation scheduling optimization model to obtain the scheduling prediction solution of the first operation scheduling optimization model, and the scheduling prediction solution is used as a scheduling configuration to perform resource scheduling on the data center energy system.

8. A training device for an end-to-end prediction optimization model for a data center energy system, characterized in that: include: A training dual construction module is used to establish a dual optimization training model of a first operation scheduling optimization training model of a data center energy system under a first operation target; A training prediction building module, used to establish an uncertain condition prediction training model of the data center energy system in a long short-term memory network according to boundary condition information corresponding to the first operation scheduling optimization training model; A sub-gradient construction module, used to determine the sub-gradient function of the decision loss function of the uncertain condition prediction training model under the dual optimization training model; A model training module is used to iteratively update the parameters of the uncertain condition prediction training model through the subgradient function to obtain a trained end-to-end prediction optimization model; the end-to-end prediction optimization model is used to determine the scheduling prediction dual solution of the dual optimization model of the first operation scheduling optimization model based on the boundary condition information corresponding to the first operation scheduling optimization model.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the method according to any one of claims 1 to 7 when executed by the processor.

Images