Method and system for algorithmic co-scheduling demand response bidding based on large model training task

Abstract

The application discloses an algorithm electricity collaborative demand response bidding decision method and system based on large model training tasks, and relates to the technical field of power system dispatching. First, the power grid invitation signal is received and analyzed to construct an invitation signal feature set. Then, the state of the intelligent algorithm center is collected, and the computing power shadow price of each large model training task is calculated; the state space and action space are constructed according to the distribution characteristics of the invitation signal feature set and the computing power shadow price, the PPO algorithm is used to solve the optimal bid, and the optimal bidding strategy is output. If the bidding is successful, the adjusted task set is selected in the order from low to high according to the computing power shadow price of each large model training task, and hierarchical control is performed as needed. Through the hierarchical response mechanism and cost consideration, when interruption must be performed, the large model training task that has just completed the "checkpoint saving" is automatically locked and suspended through the computing power shadow price sorting, the impact on the training progress is minimized, and the safety of the model parameters is ensured.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching technology, and more specifically, to a computational-coordinated demand response bidding decision-making method and system based on a large model training task. Background Technology

[0002] In the context of current artificial intelligence technology, intelligent computing centers are fundamentally different from traditional data centers (IDCs) that run stateless web services or simple batch processing jobs: large model training tasks usually involve the synchronous parallel computing of thousands of GPU cards, with high state dependence, extremely long running cycles (weeks to months), and high computing power costs per unit time.

[0003] Traditional load balancing methods often treat IT equipment as ordinary, easily disconnectable power loads. However, this simplistic and crude approach is highly susceptible to causing training cluster crashes, severe progress rollbacks, and even damage to model parameters. Once the training process of a large model is abnormally interrupted, not only is it time-consuming to reload terabytes of checkpoint parameters, but a "warm-up" phase is also required to restore full computing power. If the interruption occurs between two checkpoints, all gradient data calculated using massive amounts of electricity during that time will be lost. Summary of the Invention

[0004] In view of this, the present invention provides a computing and power collaborative demand response bidding decision method and system based on large model training tasks, which is used to solve the problems of interruption of large model training tasks in existing intelligent computing centers, easy loss of model parameters, and high recovery costs.

[0005] To achieve the above objectives, the following solution is proposed: A computational-electricity collaborative demand response bidding decision-making method based on a large model training task includes: Receive grid invitation signals from the power trading platform, analyze the grid invitation signals, and construct a high-dimensional invitation signal feature set; Collect the status of the intelligent computing center, construct a computing power shadow price model based on the status of the intelligent computing center, and calculate the computing power shadow price of training tasks for each model. Based on the feature set of invitation signals and the distribution characteristics of computing power shadow prices, a state space and action space are constructed. The PPO algorithm is used to solve for the optimal bid and output the optimal bidding strategy. If the bidding is successful, the set of tasks to be adjusted will be selected in order of increasing shadow price of computing power for each training task of the major models, and the adjustment will be controlled in stages as needed.

[0006] Preferably, the invitation signal feature set is: ; in, , These are the start and end times of the response time window, respectively. For load reduction target value or load adjustment range, This refers to the market price ceiling, real-time clearing price, or time-of-use pricing curve. Based on the benchmark compensation price, For the response event type, This is a penalty clause for breach of contract.

[0007] Preferably, the status of the intelligent computing center includes: the real-time power of each model training task, the checkpoint status, and the model convergence trend.

[0008] Preferably, the computing power shadow price model is as follows: ; in, The total economic loss due to mission interruption. Save power when tasks are interrupted.

[0009] Preferably, the total economic loss includes: checkpoint preservation cost, training progress rollback cost, and model delay opportunity cost; The total economic loss is calculated as follows: ; in, Cost of maintaining checkpoints, To cover the costs of rolling back training progress, The opportunity cost of model delay; ; in, To save time, This refers to depreciation and maintenance costs per unit time. For real-time power consumption, For real-time electricity prices; ; in, To avoid losing training time, The average power of the large model training task under computational conditions; ; in, The time value coefficient, For attenuation factor, In response to the time window, Training large models i The remaining training time, This refers to the loading time.

[0010] Preferably, the state space is: ; in, For the invitation signal feature set, The distribution characteristics of the computing power shadow price of the current task queue. This is a record of historical bidding success rates and revenue. The action space is: ; in, For the reported load reduction amount, Maximum adjustable load; In order to declare the compensation price, , These are the minimum and maximum declared compensation prices, respectively. Preferably, the process of solving for the optimal price using the PPO algorithm includes: The solution aims to maximize the net revenue of the intelligent computing center during the response period, and the reward function is:

[0011] in, This is the indicator function for the selected index. For the set of interrupted tasks, To realize the total computing power shadow price of the set of interrupted tasks with reported load reduction, Penalty for opportunity loss due to failure to win the bid.

[0012] Preferably, the process of hierarchically controlling the training tasks of various models as needed includes: If the large model training task is a stateless task, then a first-level response is performed, and traffic shaping is performed through the gateway layer to redirect some inference requests to nodes in the non-response area. If the large model training task is a non-critical path large model training task, then a second-level response is performed to execute dynamic voltage frequency adjustment and lock the GPU clock frequency at a lower frequency point with better energy efficiency. If the large model training task is of low priority or has just been completed and archived, a three-level response is performed, a checkpoint is saved, and after the data is written to disk, the video memory is released and the container is suspended.

[0013] A computing-electricity collaborative demand response bidding decision-making system based on a large model training task includes: The signal receiving and parsing module receives the grid invitation signal from the power trading platform, parses the grid invitation signal, and constructs a high-dimensional invitation signal feature set. The task value quantification module collects the status of the intelligent computing center, constructs a computing power shadow price model based on the status of the intelligent computing center, and calculates the computing power shadow price of training tasks for various models. The bidding decision module constructs a state space and action space based on the feature set of invitation signals and the distribution characteristics of computing power shadow prices, uses the PPO algorithm to solve for the optimal bid, and outputs the optimal bidding strategy. The response execution control module, upon successful bidding, selects the set of tasks to be adjusted in ascending order of the shadow price of computing power for each model training task, and controls them in a tiered manner as needed.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention provides a computational-power collaborative demand response bidding decision-making method based on large model training tasks. First, it receives power grid invitation signals from the power trading platform, analyzes these signals, and constructs a high-dimensional feature set. Then, it collects the state of the intelligent computing center, constructs a computing power shadow price model based on this state, and calculates the computing power shadow price for each training task. Based on the distribution characteristics of the invitation signal feature set and the computing power shadow price, it constructs a state space and action space, uses the PPO algorithm to solve for the optimal bid, and outputs the optimal bidding strategy. Finally, if the bid is successful, it selects the set of tasks to be adjusted according to the ascending order of the computing power shadow prices of each training task, and controls them hierarchically as needed. This invention forms a "safety barrier" through a hierarchical response mechanism and cost considerations. It prioritizes responding to power grid commands by lossless frequency reduction, maintaining memory status and communication connections, and avoiding the risk of progress rollback. If an interruption is necessary, it automatically locks and suspends the large model training task that has just completed "checkpoint saving" by sorting the computing power shadow prices. At this point, the cost of rolling back the training progress is close to zero, minimizing the impact on the training progress and ensuring the safety of the model parameters.

[0015] This invention constructs a computing power shadow pricing model to precisely quantify the implicit computing power cost of each interruption or frequency reduction operation. It combines this with a reinforcement learning bidding strategy to play a game, accepting response requests only when the grid's compensation bid covers the total economic loss, ensuring a positive return for every response. Furthermore, by offering high premiums during power shortages, the intelligent computing center is transformed from a simple electricity consumer into a virtual power plant capable of profiting from ancillary services.

[0016] This invention transforms the originally unadjustable training load into a flexible resource through quantified cost and hierarchical control. The intelligent computing center can provide the power grid with two dimensions of regulation capabilities: achieving second / millisecond-level regulation through lossless frequency reduction and rapid power reduction; and achieving minute-level regulation through graceful task interruption. These two dimensions of regulation capabilities effectively absorb curtailed wind and solar power. Attached Figure Description

[0017] 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, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart of a computing and power collaborative demand response bidding decision-making method based on a large model training task provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a computing and power collaborative demand response bidding decision-making system based on a large model training task, provided in an embodiment of the present invention. Detailed Implementation

[0019] 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.

[0020] First, combined Figure 1 This invention introduces a computing-powered collaborative demand response bidding decision-making method based on a large model training task, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes: Step S01: Receive the grid invitation signal from the power trading platform, analyze the grid invitation signal, and construct a high-dimensional invitation signal feature set.

[0021] Specifically, it connects to the power trading platform in real time to analyze grid invitation signals. The key is to extract the urgency of the demand response events (response time window). ) and incentives (benchmark compensation price) (or real-time clearing price), and automatically filter out invalid signals that are too low in incentive or have too short a response time window (i.e., unable to complete checkpoint saving).

[0022] The received power grid invitation signal The signal is parsed into a high-dimensional vector, resulting in the invitation signal feature set. :

[0023] in, , These are the start and end times of the response time window, accurate to the second, and can automatically calibrate network latency to ensure time synchronization. The target load reduction value (MW) or load adjustment range. This refers to the market price ceiling, real-time clearing price, or time-of-use electricity price curve. Depending on the event type (such as day-ahead invitation, intraday real-time scheduling, spinning reserve, frequency response), different event types will trigger different response priorities. This is a penalty clause for breach of contract, used for subsequent calculation of risk costs. The benchmark compensation price can be the historical average price or the guaranteed minimum price.

[0024] Step S02: Collect the status of the intelligent computing center, construct a computing power shadow price model based on the status of the intelligent computing center, and calculate the computing power shadow price of each training task of the major models.

[0025] Specifically, a deep scan of the intelligent computing center's status. The system iterates through the task queue. J i For each large model training task, the focus is on collecting its real-time power. (As a load base), checkpoint status (Determines the size of sunk costs) and the model convergence trend (Determine the opportunity cost of time).

[0026] Based on the collected data on the status of intelligent computing centers, a mathematical model is used to transform the physical-level disruption risk into the economic-level shadow price of computing power. (Unit: Yuan / kWh)

[0027] Define the shadow price of computing power Total economic losses caused by demand response leading to the interruption of large model training tasks With the power saved by this operation The ratio: ; in, The total economic loss due to mission interruption. Save power when tasks are interrupted.

[0028] The total economic loss consists of three parts: checkpoint preservation cost, training progress rollback cost, and model delay opportunity cost. ; in, Cost of maintaining checkpoints, To cover the costs of rolling back training progress, The opportunity cost of model delay.

[0029] The calculations for each item are as follows: (1) Cost of saving checkpoints: Model parameters must be saved before interruption.

[0030] ; in, To save time, This refers to depreciation and maintenance costs per unit time. This refers to real-time power consumption (I / O-intensive power consumption). This refers to the real-time electricity price.

[0031] (2) Training progress rollback cost: Interruption may cause the gradients of the most recent steps to be lost, requiring rollback to the previous Ckpt.

[0032] ; in, To avoid losing training time, The average power of the large model training task under computational conditions.

[0033] (3) Model delay opportunity cost: For large models, time-to-market is crucial.

[0034] ; in, Training large models J i The time value coefficient increases exponentially with the urgency of model release. The decay factor indicates that the time value is high as training nears its end. In response to the time window, Training large models i The remaining training time, This refers to the loading time required to reload checkpoints and resume training after an interruption.

[0035] When a sharp drop in the loss of a large model training task is detected, the time value coefficient of that task will be automatically increased. This causes the shadow price of computing power for the task to surge, thus protecting the task from interruption in the bidding strategy. A dynamic shadow price matrix is ​​output, clearly indicating the loss of computing power assets for every 1 kWh of electricity reduction, providing an indispensable cost floor for bidding.

[0036] Ultimately, the overall marginal response cost curve of the intelligent computing center is obtained by sorting the computing power shadow prices of all large model training tasks.

[0037] Step S03: Construct the state space and action space based on the feature set of invitation signals and the distribution characteristics of computing power shadow prices, use the PPO algorithm to solve for the optimal bid, and output the optimal bidding strategy.

[0038] Specifically, in the decision-making phase, the bidding problem is modeled as a Markov Decision Process (MDP), and the Proximal Policy Optimization (PPO) algorithm is used to solve for the optimal bid. Through offline training and online fine-tuning, the policy π(a|s) is learned. The optimal bidding strategy is then output. This means precisely determining the amount of reduction (MW) to be declared and the unit price (yuan / kWh) to be declared, with the aim of maximizing net benefits (grid compensation - computing power loss).

[0039] The state space is: ; in, For the invitation signal feature set, The distribution characteristics of the computing power shadow price of the current task queue (such as mean, variance, P90 quantile). This is a record of historical bidding success rates and revenue.

[0040] The motion space is: ; in, This refers to the declared load reduction (MW). This represents the maximum declared load reduction. The declared compensation price is (yuan / kWh). , These are the minimum and maximum declared compensation prices, respectively. The reported load reduction amount output by the decision-making agent is limited to the maximum adjustable load. And usually close to the load reduction target value published by the power grid. To optimize the direction, but not to force equality.

[0041] Reward function: The objective is to maximize the net revenue of the intelligent computing center during the response period. The reward function is as follows:

[0042] in, This is the bid-winning indicator function (1 for successful bid, 0 for unsuccessful bid). For the set of interrupted tasks, To realize the total computing power shadow price of the set of interrupted tasks with reported load reduction, Penalty for opportunity loss due to failure to win the bid.

[0043] Step S04: Determine whether the bidding was successful.

[0044] Specifically, a response will only be made if the declared compensation price is higher than the internal computing power shadow price, meaning the bid is successful.

[0045] If the bid is successful, proceed to step S05 below; if the bid is unsuccessful, wait for the next bid.

[0046] Step S05: Select the set of tasks to be adjusted in order of increasing computing power shadow price of each training task of the major models, and control them in stages as needed.

[0047] Specifically, once a bid is successful (i.e., the power grid accepts the offer), the execution phase begins. This is based on the load reduction amount requested by the winning bidder. The set of tasks to be adjusted is selected in ascending order of the shadow price of computing power for each model training task. and implement cost-aware, on-demand, tiered control: (1) Level 1 Response (Low-Cost Channel): If the large model training task is a stateless task (such as inference API, data preprocessing), then a first-level response is performed, and traffic shaping is performed through the gateway layer to redirect some inference requests to nodes in non-response areas, or temporarily reduce the concurrency of non-critical tasks.

[0048] By identifying the task ID in the HTTP request header at the gateway layer, high-priority (training data writing) and low-priority (inference preview / log query) traffic are distinguished. For low-priority tasks, a token bucket algorithm is used to limit the request rate. For example, the concurrency window size is gradually reduced in 10% increments until the real-time power of the node drops to the target threshold.

[0049] If single-node degradation cannot meet the requirements, newly arriving stateless inference requests will be redirected to a standby computing area that is not involved in demand response through a load balancer.

[0050] (2) Level II response (medium cost channel) If the large model training task is a non-critical path large model training task, a secondary response is performed to execute dynamic voltage frequency adjustment (DVFS) to achieve lossless frequency reduction. Through the management interface (such as NVIDIA NVML), the GPU clock frequency is locked at a lower frequency point with better energy efficiency, maintaining the continuity of training and avoiding the cost of checkpoint rollback.

[0051] (3) Level 3 response (high-cost channel) If the large model training task is low priority or has just been completed and archived, a three-level response is initiated: a graceful interrupt is triggered, an emergency checkpoint is saved, and after the data is written to disk, the GPU memory is released and the container is suspended. Simultaneously, the suspension state is recorded for rapid recovery after the requirement response is completed.

[0052] Next, the computing and power collaborative demand response bidding decision system based on a large model training task provided in the embodiments of the present invention will be described. The computing and power collaborative demand response bidding decision system based on a large model training task described below can be referred to in correspondence with the computing and power collaborative demand response bidding decision method based on a large model training task described above.

[0053] First, combine Figure 2 This paper introduces a computing-electricity collaborative demand response bidding decision-making system based on a large model training task, such as... Figure 2 As shown, the decision-making system may include: The signal receiving and parsing module 100 receives the power grid invitation signal from the power trading platform, parses the power grid invitation signal, and constructs a high-dimensional invitation signal feature set.

[0054] Specifically, the signal receiving and parsing module 100, acting as the system's "external sensing arm," establishes a long-term connection with the electricity market trading platform or load aggregator via an encrypted leased line (VPN / private network) or an Internet-based secure transport layer protocol (TLS 1.3). The module incorporates a multi-protocol adapter, supporting OpenADR 2.0b (Open Automatic Demand Response Standard), IEEE 2030.5 (Smart Energy Configuration Standard), and regional electricity market-customized RESTful API interfaces, ensuring compatibility with the access requirements of different electricity markets.

[0055] The Task Value Quantification Module 200 collects the status of the intelligent computing center, constructs a computing power shadow price model based on the status of the intelligent computing center, and calculates the computing power shadow price of training tasks for various models.

[0056] Specifically, the task value quantification module 200, based on the acquired multi-source data, uses a built-in shadow price calculation engine that refreshes the calculation of the current moment at a frequency of seconds (e.g., 1Hz). t The shadow price of computing power CSP(t) generated in response to DR commands (i.e., interruption or frequency reduction).

[0057] The Task Value Quantization Module 200 acts as a bridge between IT (computing power) and OT (electricity). It connects to the underlying task scheduler (such as the Slurm REST API or Kubernetes API Server) and the energy monitoring system (such as Prometheus + DCGM / IPMI) via a high-speed bus interface. It pulls real-time data from each running large model training task. J i The deep state data includes: the current loss value and its slope (to determine whether it is in a critical convergence period), the number of global training steps (GlobalStep), the physical storage timestamp of the most recent checkpoint, and the parameter size and memory usage of the current model.

[0058] The bidding decision module 300 constructs a state space and action space based on the feature set of invitation signals and the distribution characteristics of computing power shadow prices, uses the PPO algorithm to solve for the optimal bid, and outputs the optimal bidding strategy.

[0059] Specifically, the bidding decision module 300 is the "brain" of the system, with a built-in decision agent based on deep reinforcement learning (DRL). This decision agent typically uses the PPO-Clip (Proximal Policy Optimization with Clipped Objective) or SAC (Soft Actor-Critic) algorithm to find the optimal solution in a continuous action space.

[0060] The decision-making agent receives grid invitation signals (remaining response time, current electricity price, historical winning probability) from the signal receiving and parsing module 100 and internal states (shadow price distribution, adjustable task margin) from the task value quantification module 200. Unlike traditional rule engines, this decision-making agent possesses the ability to "explore and utilize," predicting market supply and demand tensions based on historical data (e.g., predicting tonight's grid shortage), and then superimposing a reasonable profit premium on the internal computing power shadow price to formulate aggressive or conservative bidding strategies, precisely determining the amount of reduction (MW) to be declared and the unit price (yuan / kWh).

[0061] The response execution control module 400, upon successful bidding, selects the set of tasks to be adjusted in ascending order of the shadow price of computing power for each model training task, and controls them in a tiered manner as needed.

[0062] Specifically, the response execution control module 400, acting as the system's "execution arm," is responsible for translating the optimal bidding strategy generated by the bidding decision module 300 (such as "reduce 5MW of load between 14:00 and 15:00") into atomic control instructions tailored to specific hardware and tasks. It supports various fine-grained operations. 1. Lossless Frequency Reduction (DVFS): By calling the GPU driver interface (such as nvidia-smi-lgc or Huawei npu-smi), the frequency of a specified cluster is locked at a lower level, achieving millisecond-level power reduction without interrupting the training process.

[0063] 2. Graceful Suspend: Sends SIGSTOP or a specific signal to the training container to trigger the system-supported "memory hold suspension" or "exit after checkpoint" to ensure that the progress is not lost.

[0064] 3. Temperature control coordination: Coordinates the adjustment of the chilled water outlet temperature of the precision air conditioner to help reduce PUE.

[0065] 4. Safety Closed-Loop Control: The module has a built-in Safety Guardrail function. Before issuing any power limiting command, it pre-verifies whether the command will cause hardware voltage instability, communication timeout (NCCL Timeout), or task crash. If a risk is detected, execution will be rejected and feedback will be sent to the bidding decision module 300 for replanning. Simultaneously, the actual power reduction during command execution will be monitored in real time to ensure that the grid's load reduction target is met. This is to avoid penalties for breach of contract due to deviations in response.

[0066] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0067] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0068] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for electricity collaborative demand response bidding decision based on large model training task, characterized in that, include: Receive grid invitation signals from the power trading platform, analyze the grid invitation signals, and construct a high-dimensional invitation signal feature set; Collect the status of the intelligent computing center, construct a computing power shadow price model based on the status of the intelligent computing center, and calculate the computing power shadow price of training tasks for each model. Based on the feature set of invitation signals and the distribution characteristics of computing power shadow prices, a state space and action space are constructed. The PPO algorithm is used to solve for the optimal bid and output the optimal bidding strategy. If the bidding is successful, the set of tasks to be adjusted will be selected in order of increasing shadow price of computing power for each training task of the major models, and the adjustment will be controlled in stages as needed.

2. The computational-electricity collaborative demand response bidding decision-making method based on a large model training task according to claim 1, characterized in that, The invitation signal feature set is as follows: ； in, , These are the start and end times of the response time window, respectively. For load reduction target value or load adjustment range, This refers to the market price ceiling, real-time clearing price, or time-of-use pricing curve. Based on the benchmark compensation price, For the response event type, This is a penalty clause for breach of contract.

3. The computational-electricity collaborative demand response bidding decision-making method based on a large model training task according to claim 1, characterized in that, The status of the intelligent computing center includes: real-time power of various model training tasks, checkpoint status, and model convergence trend.

4. The computational-electricity collaborative demand response bidding decision-making method based on a large model training task according to claim 1, characterized in that, The computing power shadow price model is as follows: ； in, The total economic loss due to mission interruption. Save power when tasks are interrupted.

5. The computational-electricity collaborative demand response bidding decision-making method based on a large model training task according to claim 4, characterized in that, The total economic loss includes: checkpoint preservation cost, training progress rollback cost, and model delay opportunity cost; The total economic loss is calculated as follows: ； in, Cost of maintaining checkpoints, To cover the costs of rolling back training progress, The opportunity cost of model delay; ； in, To save time, This refers to depreciation and maintenance costs per unit time. For real-time power consumption, For real-time electricity prices; ； in, To avoid losing training time, The average power of the large model training task under computational conditions; ； in, The time value coefficient, For attenuation factor, In response to the time window, Training large models i The remaining training time, This refers to the loading time.

6. The computational-electricity collaborative demand response bidding decision-making method based on a large model training task according to claim 1, characterized in that, The state space is as follows: ； in, For the invitation signal feature set, The distribution characteristics of the computing power shadow price of the current task queue. This is a record of historical bidding success rates and revenue. The action space is: ； in, For the reported load reduction amount, Maximum adjustable load; In order to declare the compensation price, , These are the minimum and maximum declared compensation prices, respectively.

7. The computational-electricity collaborative demand response bidding decision-making method based on a large model training task according to claim 6, characterized in that, The process of finding the optimal price using the PPO algorithm includes: The solution aims to maximize the net revenue of the intelligent computing center during the response period, and the reward function is: in, This is the indicator function for the selected index. For the set of interrupted tasks, To realize the total computing power shadow price of the set of interrupted tasks with reported load reduction, Penalty for opportunity loss due to failure to win the bid.

8. The computational-electricity collaborative demand response bidding decision-making method based on a large model training task according to claim 6, characterized in that, The process of hierarchical control over the training tasks of various models on demand includes: If the large model training task is a stateless task, then a first-level response is performed, and traffic shaping is performed through the gateway layer to redirect some inference requests to nodes in the non-response area. If the large model training task is a non-critical path large model training task, then a second-level response is performed to execute dynamic voltage frequency adjustment and lock the GPU clock frequency at a lower frequency point with better energy efficiency. If the large model training task is of low priority or has just been completed and archived, a three-level response is performed, a checkpoint is saved, and after the data is written to disk, the video memory is released and the container is suspended.

9. A computing-electricity collaborative demand response bidding decision-making system based on a large model training task, characterized in that, include: The signal receiving and parsing module receives the grid invitation signal from the power trading platform, parses the grid invitation signal, and constructs a high-dimensional invitation signal feature set. The task value quantification module collects the status of the intelligent computing center, constructs a computing power shadow price model based on the status of the intelligent computing center, and calculates the computing power shadow price of training tasks for various models. The bidding decision module constructs a state space and action space based on the feature set of invitation signals and the distribution characteristics of computing power shadow prices, uses the PPO algorithm to solve for the optimal bid, and outputs the optimal bidding strategy. The response execution control module, upon successful bidding, selects the set of tasks to be adjusted in ascending order of the shadow price of computing power for each model training task, and controls them in a tiered manner as needed.

Images