A conformal calibration electricity price prediction method for computerized collaborative decision-making

Abstract

The application discloses a conformal calibration electricity price prediction method for computer-aided collaborative decision-making, and belongs to the field of power distribution network optimization; the method is applied to a collaborative system containing flexible resources on the computing power side and the power side, and comprises the following steps: in a training stage, a neural network is trained by using historical characteristic information, a conformal threshold is determined based on a conformal prediction theory, and a historical electricity price box type uncertainty set is constructed; the set is input into a differentiable convex optimization layer which is reconstructed from a robust scheduling model dual, and a total actual operation cost is taken as a loss function to perform decision focusing back propagation training on the prediction network. In a day-ahead scheduling stage, a trained prediction network and the conformal threshold are used to generate a calibrated day-ahead electricity price box type uncertainty set, and the set is input into the differentiable convex optimization layer again to obtain optimal scheduling decision variables containing the states of the two types of flexible resources and to issue control instructions. The application reduces the comprehensive operation cost of the system under the premise of guaranteeing uncertainty statistical coverage.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network optimization, and specifically relates to a conformal calibration electricity price prediction method for computer-based collaborative decision-making. Background Technology

[0002] As the scale of artificial intelligence tasks continues to grow, the electricity demand of data centers is rising rapidly. Coordinated optimization between power-side and computing-side resources in the distribution network is gradually becoming a key technological direction for improving system economy and sustainability. Currently, in distribution networks connected to distributed photovoltaic systems, energy storage devices, and data centers with load migration capabilities, the dispatching entity typically needs to coordinate power purchases, data center workload migration, energy storage charging and discharging, and local adjustable power output based on day-ahead electricity price forecasts. In existing technologies, most electricity price forecasting models only aim to minimize prediction errors, while dispatching models aim to minimize operating costs. These two are independent, forming a separate technical path of "prediction first, optimization later."

[0003] However, electricity price forecasting errors do not affect dispatching results uniformly. In power-computing collaboration scenarios, deviations at certain times can significantly alter the main grid's power purchase capacity, data center load migration paths, energy storage charging and discharging status, and renewable energy consumption methods, leading to increased actual operating costs. While existing uncertainty handling techniques can provide interval or probabilistic forecasting results, they often lack targeted designs for downstream dispatching tasks, making it difficult to simultaneously consider statistical coverage, dispatch solvability, and operational economics. Especially in scenarios involving the joint operation of distribution networks, data centers, and renewable energy, how to construct an uncertainty set that satisfies both statistical reliability and can directly serve robust dispatching solutions has become a crucial issue in existing technologies.

[0004] Existing technologies in power-computing collaborative scheduling suffer from at least the following shortcomings: First, the electricity price forecasting stage and the scheduling decision-making stage are isolated from each other, and the forecasting model does not consider downstream operating costs, so improved forecasting accuracy does not necessarily lead to improved scheduling performance. Second, traditional interval forecasting results lack coverage calibration, making it difficult to provide a stable and reliable uncertainty representation for unknown samples. Third, there is insufficient coupling between the uncertainty set and the robust optimization model, making it difficult to balance computational tractability and scheduling conservatism. Fourth, existing methods fail to uniformly consider the collaborative relationship between data center load migration, energy storage regulation, distributed power sources, and distribution network constraints. All of these problems urgently need to be addressed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a conformal calibration electricity price prediction method for computer collaborative decision-making, thereby solving the problems in existing technologies.

[0006] The objective of this invention can be achieved through the following technical solutions: A conformal calibration electricity price prediction method for collaborative decision-making in computing systems, applied to power-computing collaborative systems, includes: Obtain historical contextual features of the power-computing power collaborative system and the corresponding real electricity price; A neural network predictor is trained using historical contextual feature information to output an electricity price prediction interval; the non-consistency score of the actual electricity price to the electricity price prediction interval is calculated using a calibration dataset, and a conformal threshold is determined based on the non-consistency score. The electricity price prediction interval is then expanded using the conformal threshold to construct a box-shaped uncertainty set of historical electricity prices. The uncertainty set of the electricity price box type is input into a preset differentiable convex optimization layer to obtain the optimal scheduling decision variables. The differentiable convex optimization layer is reconstructed from a robust scheduling model that includes computing power side constraints and power side constraints. Using the actual total operating cost of the optimal scheduling decision variable under the real electricity price as the task loss function, the gradient of the task loss function with respect to the parameters of the neural network predictor is obtained by taking the derivative, and backpropagation is used to update the parameters to obtain the trained neural network predictor. The contextual feature information of the current scheduling cycle is obtained and input into the trained neural network predictor to obtain the day-ahead electricity price prediction interval. Based on the conformal threshold, the prediction interval is calibrated and expanded to generate the day-ahead electricity price box-shaped uncertainty set. The optimal scheduling decision variable is solved based on the day-ahead electricity price box-shaped uncertainty set, and the optimal scheduling decision variable is used as the control basis for power-computing power collaborative scheduling.

[0007] Furthermore, the power-computing collaborative system includes: The physical layer includes a power distribution network, multiple geographically distributed data centers, distributed photovoltaics, battery energy storage systems, and a main grid power purchase interface. The data centers are configured to provide computing power flexibility through workload migration and to provide power-side flexibility using the battery energy storage systems, so as to achieve coordinated and optimized scheduling of power resources and computing power resources. The information layer includes a neural network prediction module and a collaborative optimization decision module. The neural network prediction module is used to generate day-ahead electricity price prediction intervals and construct a box-shaped uncertainty set of electricity prices. The collaborative optimization decision module is used to determine the workload migration strategy between data centers, the charging and discharging power of the battery energy storage system, and the power purchased by the main grid based on the box-shaped uncertainty set of electricity prices.

[0008] Furthermore, the contextual feature information includes at least: historical electricity price sequences, weather conditions, load forecast results, and calendar indicators. The neural network predictor is configured to receive the contextual feature information and output the lower bound and upper bound predicted values ​​of electricity prices for each time period within the scheduling cycle.

[0009] Furthermore, the robust scheduling model takes minimizing the total operating cost of the power-computing collaborative system in the worst case as its objective function. The total operating cost of the power-computing collaborative system includes: diesel generator generation cost, diesel generator start-up cost, main grid power purchase cost, and data center workload migration cost. The computing power-side constraints include data center load migration model constraints; the power-side constraints include battery energy storage model constraints, diesel generator set model constraints, and power distribution network operation model constraints.

[0010] Furthermore, the conformal threshold determination process includes: Define the inconsistency scoring function as follows: Where x represents contextual feature information, For the true electricity price vector, and These are the predicted lower bound vector and the predicted upper bound vector, respectively; based on a calibration dataset containing N samples, the inconsistency score for each sample is calculated, and empirical values ​​are taken. The quantile is used as the shape-preserving threshold q.

[0011] Furthermore, the uncertainty set of the electricity price box type is: .

[0012] Furthermore, the reconstruction process of the differentiable convex optimization layer includes: The inner maximization problem regarding uncertain electricity prices in the robust scheduling model is dualized by introducing a non-negative auxiliary variable v. The cost item containing the uncertain electricity price, the deterministic operating cost item, and the set of physical boundary constraints are reconstructed into a differentiable objective function, the expression of which at least includes: ,in, This is the lower limit of the electricity price. Let z be the upper bound of the electricity price, and z be the dispatch decision variable. This represents the mapping relationship between the power interacting with the external power grid and the scheduling decision variables.

[0013] Furthermore, the task loss function is: In the formula, Indicates in the uncertain set The optimal scheduling decision is obtained by finding the optimal scheduling decision. The gradient of the task loss function with respect to the model parameters is obtained by differentiating the Karush-Kuhn-Tucker condition of the differentiable convex optimization layer.

[0014] A power-computing power collaborative scheduling device based on decision-focused conformal calibration executes the above method, including: Data acquisition module: Acquires historical contextual feature information and corresponding real electricity prices from the power-computing collaborative system; Uncertainty set construction module: Train a neural network predictor using historical contextual feature information to output the electricity price prediction interval; Calculate the non-consistency score of the actual electricity price to the electricity price prediction interval using a calibration dataset, determine the conformal threshold based on the non-consistency score, and expand the electricity price prediction interval using the conformal threshold to construct the box-shaped uncertainty set of electricity price. Scheduling decision variable acquisition module: Input the uncertain set of the electricity price box type into a preset differentiable convex optimization layer to obtain the optimal scheduling decision variables. The differentiable convex optimization layer is reconstructed from a robust scheduling model that includes computing power side constraints and power side constraints. Model training module: Using the actual total operating cost under the historical real electricity price as the historical optimal scheduling decision variable as the task loss function, the gradient of the task loss function with respect to the parameters of the neural network predictor is obtained by taking the derivative, and backpropagation is used to update the parameters to obtain the trained neural network predictor.

[0015] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to perform the steps of the method described above.

[0016] The beneficial effects of this invention are: 1. This invention employs an end-to-end training mechanism of Decision Focused Learning (DFL), directly using the actual worst-case scheduling cost as the task loss function. By differentiating the reconstructed differentiable convex optimization layer, the gradient is backpropagated to the neural network predictor for model parameter updates. This breaks down the barriers between electricity price prediction and downstream scheduling, enabling the prediction model to perceive the impact of prediction deviations on the actual downstream scheduling costs, rather than simply pursuing statistical error indicators. Compared to the traditional fragmented approach of prediction followed by optimization, this method effectively reduces the actual overall operating cost of the power-computing collaborative system when facing day-ahead electricity price fluctuations.

[0017] 2. This invention introduces conformal prediction theory. After obtaining the initial prediction interval, it calculates the non-consistency score of the actual electricity price based on the calibration dataset and selects the empirical quantile as the conformal threshold to dynamically expand and calibrate the initial interval, thus constructing a box-shaped uncertainty set of the electricity price. This changes the traditional approach of relying on human experience to set the proportion or assume the probability distribution for the uncertainty set. Under a given risk level, it provides a statistically significant guarantee of coverage for the day-ahead electricity price interval, thereby improving the reliability and stability of the robust dispatch scheme in the face of real electricity price scenarios.

[0018] 3. This invention addresses robust scheduling models with uncertain electricity prices by introducing non-negative auxiliary variables to dualize the inner-layer maximization problem. This transforms the complex two-stage minimization-maximization robust optimization problem into a single-layer differentiable convex optimization model. This not only provides a differentiable mathematical basis for subsequent gradient backpropagation in neural networks but also significantly reduces the computational complexity of solving constrained robust scheduling problems, improving the convergence and computability of the algorithm in practical power distribution network and data center scheduling applications.

[0019] 4. This invention constructs a joint constraint set of multi-source heterogeneous physical boundaries, including power flow in the distribution network, distributed photovoltaics, mutual exclusion of energy storage charging and discharging, data center computing power load migration conservation, and network link capacity. It establishes a clear mathematical mapping between computing power transfer and the active power of electrical nodes. By incorporating the network flexibility on the computing power side and the energy storage / generator regulation capability on the power side into a unified physical feasible domain for global coordination, it promotes cross-boundary interaction of flexible regulation resources, improves the utilization rate of data center infrastructure, and enhances the distribution network's ability to absorb distributed renewable energy to a certain extent. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the power-computing power collaborative scheduling method of the present invention; Figure 2 This is a framework diagram of the power-computing power collaborative scheduling system of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 A conformal calibration method for electricity price forecasting oriented towards computer-aided collaborative decision-making is proposed. This method uses conformal prediction (CP) to calibrate day-ahead electricity price uncertainty and employs decision-focused learning (DFL) to couple the uncertainty set generation process with the downstream robust scheduling process end-to-end. This ensures that the generated day-ahead electricity price uncertainty set simultaneously considers statistical coverage reliability and downstream decision adaptability, making it suitable for electricity price uncertainty forecasting in scenarios involving the coordinated operation of distribution networks, data centers, and renewable energy sources. Figure 1 As shown, the specific implementation steps are as follows: (1) Construct a power-computing power collaborative scheduling system framework; In this step, such as Figure 2 As shown, a collaborative operation framework is established, encompassing a distribution network, multiple geographically distributed data centers, distributed photovoltaic systems, battery energy storage systems, and a mains grid power purchase interface. The distribution network operator forecasts future day-ahead electricity prices based on contextual characteristics and accordingly coordinates workload migration between data centers, energy storage charging and discharging power, local diesel generator output, and mains grid power purchase. These contextual characteristics include at least historical electricity price sequences, weather conditions, load forecasting results, and calendar indicators. Data centers provide computing power flexibility through workload migration and power flexibility through local battery energy storage systems, thereby achieving coordinated regulation of power and computing resources.

[0024] (2) Construct a power-computing power collaborative scheduling and operation model; To make the downstream scheduling problem computable, this invention establishes a data center load migration model, a battery energy storage model, a diesel generator set model, and a power distribution network operation model, respectively.

[0025] Specifically, for data center load migration, the net migration load of data center i in time period t is defined as: The load transferred from data center i to data center j is Then, the following conditions must be met: the migration net load of a data center equals the difference between its outgoing workload and its incoming workload; the sum of the migration net loads of all data centers is zero to ensure the conservation of total workload; the data center load for each time period is recursively derived from the baseline load and the migration net load; the migration net load is affected by the flexibility factor. The workload transfer volume is limited by the link capacity. limit.

[0026] In the formula, This represents the change in data center net load caused by workload migration; This represents the workload that is migrated from data center i to data center j; This indicates the proportion of data center i that can be migrated; This represents the maximum transferable capacity between data center i and data center j.

[0027] For battery energy storage systems, the state of charge of data center i in time period t is defined as follows: Charging power is The discharge power is If the energy storage state satisfies the initial state of charge constraint, the recursive state of charge constraint, the upper and lower limit constraints, and the mutual exclusion constraint that "charging and discharging cannot occur simultaneously".

[0028] In the formula, Indicates charging efficiency; Indicates discharge efficiency; This represents a binary variable indicating the energy storage's operating status. The net active power presented by the data center to the distribution network is... In the formula, This represents the net active power demand at the data center i's access point to the power distribution network node.

[0029] For the diesel generator set model, the power generation of node i in time period t is defined as follows: The unit's start / stop status is Therefore, the unit output is subject to upper and lower limits as well as ramping constraints, and starting variables are introduced. Startup costs This is used to characterize the start-up and shutdown costs of the generating unit.

[0030] In the formula, and These represent the upper and lower limits of the unit's ramp rate, respectively; This represents the unit startup cost coefficient.

[0031] For the distribution network operation model, a linearized power flow model is used to describe the active power balance, reactive power balance, voltage drop, and upper and lower limit constraints of node voltage at each node, while also considering the upper limit constraint of distributed photovoltaic power output.

[0032] In the formula, This represents the electrical power purchased from the main grid during time period t; This indicates the active power output of the photovoltaic system at node i; This represents the voltage amplitude of node i during time period t.

[0033] Based on the above model, a day-ahead robust cooperative scheduling objective function is constructed to minimize the overall system operating cost under the worst-case electricity price scenario. The overall operating cost includes at least the cost of diesel generator generation, diesel generator start-up cost, mains grid power purchase cost, workload migration cost, and workload migration delay penalty cost. The objective function can be expressed as: in, Represents the electricity price vector. This represents the joint decision variables of workload migration, energy storage charging and discharging, diesel engine output, main grid power purchase, and distribution network operating status. This represents the cost of migrating a unit workload between data centers. This represents the unit workload migration delay penalty coefficient caused by data center inter-data center transmission distance, network communication latency, or service level agreement constraints.

[0034] (3) Construct a conformal calibration uncertainty set for day-ahead electricity prices; In this step, we first construct a neural network predictor. To predict the lower and upper bounds of the day-ahead electricity price vector based on contextual feature x, i.e. In the formula, Indicates the parameters of the prediction model; This represents the lower bound vector for the predicted electricity price; This represents the upper bound vector of the predicted electricity price.

[0035] The neural network predictor employs a multi-layer feedforward neural network structure, including an input layer, multiple hidden layers, and an output layer. The hidden layers consist of fully connected layers and a non-linear activation function, used to extract the non-linear mapping relationship between contextual features and day-ahead electricity prices. Each hidden layer uses a three-layer fully connected structure with 256 neurons in each layer, and the activation function is the Modified Linear Unit (ReLU). The output layer has an output dimension of 2^n, where the first n dimensions represent the lower bound of the day-ahead electricity price prediction, and the last n dimensions represent the difference between the upper and lower bounds, ensuring that the upper bound of the prediction is not lower than the lower bound.

[0036] Subsequently, the inconsistent scoring function is defined. Used to measure real electricity prices The degree of violation of the prediction interval. Where, When the value is less than or equal to zero, it means that the actual electricity price falls within the predicted range; A value greater than zero indicates that the actual electricity price exceeds the predicted range.

[0037] Based on calibration dataset To accommodate the non-stationary nature of day-ahead electricity price time series, the calibration dataset... A sliding window approach is used for dynamic updating. Specifically, at the end of each scheduling cycle, the latest obtained real electricity price sample and its corresponding contextual features are added to the calibration dataset, while the oldest historical samples are removed to maintain the stability of the calibration dataset size. By dynamically updating the calibration dataset, the inconsistency score can continuously reflect the current electricity price distribution characteristics, thereby improving the adaptability of the conformal threshold to electricity price time-series drift and enhancing the coverage reliability of the prediction interval in continuous operation scenarios. Based on dynamic updating... Calculate the inconsistency score for each sample separately, and take its empirical value. Quantiles are used as the shape-preserving threshold q, i.e. In the formula, q represents the risk level; N represents the number of calibration samples; and q represents the calibration threshold used to expand the prediction interval.

[0038] The prediction interval is expanded using the conformal threshold q to construct the day-ahead electricity price box-type uncertainty set. In the formula, Let x represent the day-ahead electricity price uncertainty set corresponding to the context feature x. Through the above processing, the resulting uncertainty set has statistical coverage guarantees at a given risk level, while maintaining a box-shaped structure to facilitate coupling with downstream robust optimization models.

[0039] (4) Construct a decision-focused robust optimization and differentiable training mechanism; The box-shaped uncertainty set Substituting into the day-ahead robust cooperative scheduling model, we obtain a mini-maximum optimization problem for the scheduling variable z. Since the uncertain electricity price only affects the main grid's electricity purchase cost, the robust optimization problem can be divided into the sum of the deterministic cost term and the uncertain electricity price term. Furthermore, the inner-layer maximization problem is dualized, resulting in an equivalent resolvable model.

[0040] Specifically, let the lower bound of the box-shaped uncertain set be... The upper boundary is Introducing an auxiliary variable v, the dualized robust optimization objective can be written as: and satisfy , and various operational constraints In the formula, Fz represents the mapping relationship between the main grid purchased power and the dispatch variable z; Represents deterministic operating costs independent of uncertain electricity prices; This represents the set of constraints for comprehensive operation. Through dualization, the original robust problem can be transformed into a differentiable convex optimization layer.

[0041] During the training phase, a two-stage training mechanism is employed to optimize the parameters of the neural network predictor. First, in the pre-training phase, historical real electricity price tags are used for supervised training of the neural network predictor. A quantile loss function is used as the supervised loss to initialize the model parameters, enabling the neural network predictor to possess basic day-ahead electricity price range prediction capabilities. The neural network predictor outputs a lower bound and an upper bound for the day-ahead electricity price prediction, and is trained on the corresponding quantiles using a quantile loss function, the expression of which is: In the formula, Represents the actual day-ahead electricity price; This indicates the lower bound of the predicted electricity price. This indicates the upper limit of the predicted electricity price. Indicates the number of training samples. This represents the number of scheduling periods. After pre-training, the process enters the decision-oriented fine-tuning phase. In this phase, the actual comprehensive operating cost under the influence of real electricity prices is used as the task loss function, i.e.: In the formula, Indicates in the uncertain set The optimal scheduling decision is obtained by further optimizing the parameters, with the objective of minimizing the average task loss on the training sample set. To update, i.e. minimize During the training phase, to ensure the differentiability of the optimization layer, continuous relaxation is applied to the binary discrete decision variables in the model when solving the robust scheduling model and performing gradient backpropagation. The original 0-1 integer variables are relaxed to continuous variables within the interval [0,1], thus transforming the original mixed integer optimization problem into a differentiable continuous optimization problem. By differentiating the Karush-Kuhn-Tucker conditions of the optimization layer, the gradient of the optimal decision with respect to the model parameters is obtained, and this gradient is backpropagated to the neural network predictor. This allows the prediction model to no longer merely pursue statistically significant prediction accuracy, but instead directly learn uncertain sets that are beneficial for reducing actual scheduling costs.

[0042] (5) Perform day-ahead power-computing collaborative scheduling; In practical applications, the process first inputs contextual features such as historical electricity prices, weather, load forecasts, and calendar indicators, which are then processed by a neural network predictor to output initial upper and lower bounds for day-ahead electricity prices. Next, a conformal threshold is calculated based on a calibration set, forming a box-shaped uncertainty set. Then, this uncertainty set is input into a robust collaborative scheduling model to obtain workload migration schemes, energy storage charging and discharging power, diesel engine output, and mains grid power purchases. Finally, the day-ahead operation plan is executed according to the optimal scheduling results. This process improves the economy and stability of collaborative scheduling between the power supply side and the computing power side while ensuring statistical coverage.

[0043] Example 2 A power-computing collaborative scheduling device based on decision-focused conformal calibration specifically includes: Data acquisition module: Acquires historical contextual feature information and corresponding historical real electricity prices from the power-computing collaborative system; Uncertainty set construction module: Train a neural network predictor using historical contextual feature information to output historical electricity price prediction interval; Calculate the inconsistency score of historical real electricity price on the historical electricity price prediction interval using calibration dataset, determine the conformal threshold based on the inconsistency score, and expand the historical electricity price prediction interval using the conformal threshold to construct a box-shaped uncertainty set of historical electricity price. Scheduling decision variable acquisition module: Input the historical electricity price box-type uncertainty set into a preset differentiable convex optimization layer to obtain the historical optimal scheduling decision variables. The differentiable convex optimization layer is reconstructed from a robust scheduling model that includes computing power side constraints and power side constraints. Model training module: Using the actual total operating cost under the historical real electricity price as the historical optimal scheduling decision variable as the task loss function, the gradient of the task loss function with respect to the parameters of the neural network predictor is obtained by taking the derivative, and backpropagation is used to update the parameters to obtain the trained neural network predictor. The optimized scheduling module obtains the contextual feature information of the current scheduling cycle, inputs it into the trained neural network predictor, obtains the day-ahead electricity price prediction interval, and performs calibration expansion based on the conformal threshold to generate the day-ahead electricity price box-shaped uncertainty set, and inputs it into the differentiable convex optimization layer to obtain the optimal scheduling decision variables; finally, it issues control commands according to the optimal scheduling decision variables to execute power-computing collaborative scheduling.

[0044] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A conformal calibration electricity price prediction method for collaborative decision-making in computing, applied to a power-computing collaborative system, characterized in that, include: Obtain historical contextual features of the power-computing power collaborative system and the corresponding real electricity price; A neural network predictor is trained using historical contextual features to output an electricity price prediction range. The non-consistency score of the actual electricity price to the electricity price prediction interval is calculated using the calibration dataset, and a conformal threshold is determined based on the non-consistency score. The conformal threshold is then used to expand the electricity price prediction interval to construct a box-shaped uncertainty set of historical electricity prices. The uncertainty set of the electricity price box type is input into a preset differentiable convex optimization layer to obtain the optimal scheduling decision variables. The differentiable convex optimization layer is reconstructed from a robust scheduling model that includes computing power side constraints and power side constraints. Using the actual total operating cost of the optimal scheduling decision variable under the real electricity price as the task loss function, the gradient of the task loss function with respect to the parameters of the neural network predictor is obtained by taking the derivative, and backpropagation is used to update the parameters to obtain the trained neural network predictor. The contextual feature information of the current scheduling cycle is obtained and input into the trained neural network predictor to obtain the day-ahead electricity price prediction interval. Based on the conformal threshold, the prediction interval is calibrated and expanded to generate the day-ahead electricity price box-shaped uncertainty set. The optimal scheduling decision variable is solved based on the day-ahead electricity price box-shaped uncertainty set, and the optimal scheduling decision variable is used as the control basis for power-computing power collaborative scheduling.

2. The conformal calibration electricity price prediction method for computer-based collaborative decision-making according to claim 1, characterized in that, The power-computing collaborative system includes: The physical layer includes a power distribution network, multiple geographically distributed data centers, distributed photovoltaics, battery energy storage systems, and a main grid power purchase interface. The data centers are configured to provide computing power flexibility through workload migration and to provide power-side flexibility using the battery energy storage systems, so as to achieve coordinated and optimized scheduling of power resources and computing power resources. The information layer includes a neural network prediction module and a collaborative optimization decision module. The neural network prediction module is used to generate day-ahead electricity price prediction intervals and construct a box-shaped uncertainty set of electricity prices. The collaborative optimization decision module is used to determine the workload migration strategy between data centers, the charging and discharging power of the battery energy storage system, and the power purchased by the main grid based on the box-shaped uncertainty set of electricity prices.

3. The conformal calibration electricity price prediction method for computer-based collaborative decision-making according to claim 1, characterized in that, The contextual feature information includes at least: historical electricity price sequence, weather conditions, load forecast results, and calendar indicators. The neural network predictor is configured to receive the contextual feature information and output the lower bound and upper bound predicted values ​​of electricity prices for each time period within the scheduling cycle.

4. The conformal calibration electricity price prediction method for computer-based collaborative decision-making according to claim 1, characterized in that, The robust scheduling model takes minimizing the total operating cost of the power-computing collaborative system in the worst case as its objective function. The total operating cost of the power-computing collaborative system includes: diesel generator generation cost, diesel generator start-up cost, main grid power purchase cost, and data center workload migration cost. The computing power-side constraints include data center load migration model constraints; the power-side constraints include battery energy storage model constraints, diesel generator set model constraints, and power distribution network operation model constraints.

5. The conformal calibration electricity price prediction method for computer-based collaborative decision-making according to claim 1, characterized in that, The conformal threshold determination process includes: Define the inconsistency scoring function as follows: Where x represents contextual feature information, For the true electricity price vector, and These are the predicted lower bound vector and the predicted upper bound vector, respectively; based on a calibration dataset containing N samples, the inconsistency score for each sample is calculated, and empirical values ​​are taken. The quantile is used as the shape-preserving threshold q.

6. The conformal calibration electricity price prediction method for computer-based collaborative decision-making according to claim 5, characterized in that, The uncertainty set of the electricity price box type is: .

7. A conformal calibration electricity price prediction method for computer-based collaborative decision-making as described in claim 1, characterized in that, The reconstruction process of the differentiable convex optimization layer includes: The inner maximization problem regarding uncertain electricity prices in the robust scheduling model is dualized by introducing a non-negative auxiliary variable v. The cost item containing the uncertain electricity price, the deterministic operating cost item, and the set of physical boundary constraints are reconstructed into a differentiable objective function, the expression of which at least includes: ,in, This is the lower limit of the electricity price. Let z be the upper bound of the electricity price, and z be the dispatch decision variable. This represents the mapping relationship between the power interacting with the external power grid and the scheduling decision variables.

8. A conformal calibration electricity price prediction method for computer-based collaborative decision-making according to claim 6, characterized in that, The task loss function is: In the formula, Indicates in the uncertain set The optimal scheduling decision is obtained by finding the optimal scheduling decision. The gradient of the task loss function with respect to the model parameters is obtained by differentiating the Karush-Kuhn-Tucker condition of the differentiable convex optimization layer.

9. A power-computing power collaborative scheduling device based on decision-focused conformal calibration, comprising the method described in any one of claims 1-8, characterized in that, include: Data acquisition module: Acquires historical contextual feature information and corresponding real electricity prices from the power-computing collaborative system; Uncertain set construction module: Uses historical contextual feature information to train a neural network predictor and outputs the electricity price prediction range; The non-consistency score of the actual electricity price to the electricity price prediction interval is calculated using the calibration dataset, and a conformal threshold is determined based on the non-consistency score. The conformal threshold is then used to expand the electricity price prediction interval to construct a box-shaped uncertainty set for electricity prices. Scheduling decision variable acquisition module: Input the uncertain set of the electricity price box type into a preset differentiable convex optimization layer to obtain the optimal scheduling decision variables. The differentiable convex optimization layer is reconstructed from a robust scheduling model that includes computing power side constraints and power side constraints. Model training module: Using the actual total operating cost under the historical real electricity price as the historical optimal scheduling decision variable as the task loss function, the gradient of the task loss function with respect to the parameters of the neural network predictor is obtained by taking the derivative, and backpropagation is used to update the parameters to obtain the trained neural network predictor.

10. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the steps of the method as described in any one of claims 1-8.

Images