Methods, devices, equipment, media, and products for collaborative scheduling of computing power and power.

By combining the AAT model and the STGCN network, generative adversarial networks and reinforcement learning are used to optimize computing power and power dispatch, solving the problem of intermittent renewable energy generation and its integration with traditional power. This achieves precise resource allocation and energy consumption optimization, reduces operating costs, and improves resource utilization efficiency.

CN123114716APending Publication Date: 2026-11-17CHINA MOBILE GROUP ZHEJIANG +3
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510624832.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the existing computing-based power dispatch process, the intermittency and volatility of new energy power generation are difficult to integrate with the stable supply characteristics of traditional power, resulting in difficulties in unified dispatch, high operation and maintenance costs, obstacles to data collection and sharing, energy consumption control problems, and difficulty in achieving precise matching.

Method used

By combining the AAT model and the STGCN network, the power demand of the computing hub and the power generation of the power supply point are predicted. The accurate power allocation scheme is generated through the power allocation scheme prediction model, and the scheduling decision is optimized by using generative adversarial networks and reinforcement learning.

Benefits of technology

It enables accurate forecasting of electricity in different regions and of different types, optimizes power generation plans, reduces operating costs, improves the capacity for renewable energy absorption, and enhances resource utilization efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114716A_ABST
    Figure CN123114716A_ABST
Patent Text Reader

Abstract

This invention provides a method, apparatus, device, medium, and product for collaborative scheduling of computing power and electricity. The method includes: acquiring historical electricity consumption data and historical power generation data from a computing power network; inputting the historical electricity consumption data into a pre-trained AAT model to obtain electricity consumption prediction results; inputting the historical power generation data into a pre-trained STGCN network to obtain power generation prediction results; and inputting the electricity consumption prediction results and power generation prediction results into a pre-trained computing power scheme prediction model to obtain a computing power allocation scheme. This invention utilizes the AAT model to mine long-term dependencies and complex patterns in computing power data, thereby understanding the computing power usage patterns in different regions. Based on the Spatiotemporal Graph Convolutional Network (STGCN) algorithm, it comprehensively and accurately predicts the power generation patterns of different regions and different types of electricity, providing effective data support for collaborative scheduling of computing power and improving the resource utilization efficiency of the computing power network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of computing power and power grid integration technology, and in particular to a method, apparatus, equipment, medium and product for collaborative scheduling of computing power and power. Background Technology

[0002] Computing power-electricity coordinated dispatch refers to the dynamic matching and bidirectional optimization of computing resources (such as data centers) with power planning, operation and market levels through technological and mechanism innovation. Its core objective is to improve the efficiency of power resource utilization, promote the consumption of new energy sources (such as hydrogen energy and solar energy), save energy and reduce emissions, and enhance grid stability while ensuring the quality of computing power services.

[0003] However, existing power dispatching processes face numerous challenges. First, traditional thermal and hydropower generation is relatively stable and can be precisely controlled based on demand. In contrast, new energy power, such as photovoltaics which depends on sunlight intensity and wind power which depends on wind speed, exhibits intermittent and fluctuating generation characteristics, with varying output curves. This makes it difficult to integrate with the stable supply characteristics of traditional power, posing a significant challenge to unified dispatching.

[0004] Secondly, the standards for grid connection and management differ: Traditional power has mature grid connection specifications, equipment operation and management standards, and safety assurance systems, and its operation and maintenance are mostly centralized. The operation and maintenance management model of new energy power generation equipment is very different from that of traditional power. Due to the dispersed nature of the equipment and its remote location, the operation and maintenance costs are high and the timeliness is poor, which increases the difficulty of centralized management and may even affect the stability of the power grid.

[0005] Secondly, there are obstacles to data acquisition and sharing: In traditional power systems, data acquisition is relatively centralized and standardized, with uniform data formats, facilitating transmission and analysis. New energy power generation equipment comes from a wide range of sources, data acquisition points are scattered, sensor types vary, resulting in chaotic data formats and poor data compatibility between different manufacturers' equipment. This makes efficient data sharing and integration difficult, hindering comprehensive control over the power system's operational status.

[0006] Finally, there is the challenge of energy consumption management: computing power operation consumes enormous amounts of energy, which dynamically changes with the computing load. Currently, there is a lack of an effective computing power-power coordinated energy consumption management mechanism, making it difficult to dynamically adjust the allocation of computing power tasks according to power supply, and accurately measuring and predicting energy consumption is extremely difficult. When centrally managed, it is urgent to solve the problem of meeting computing power needs while taking into account power supply constraints, achieving energy consumption optimization, and avoiding power waste or insufficient power supply due to excessive energy consumption.

[0007] In summary, the field of computing power-electricity collaborative scheduling faces a complex challenge: the different power characteristics, operation and maintenance methods, and data formats of new energy and traditional energy make integration difficult; and there is a mismatch between the supply and demand of power resources and computing power needs. Summary of the Invention

[0008] This invention provides a method, apparatus, equipment, medium, and product for the coordinated scheduling of computing power and electricity, which addresses the deficiency in the prior art where supply and demand are difficult to match precisely, enabling comprehensive and accurate prediction of power generation patterns in different regions and of different types, and providing precise scheduling of electricity of different energy types.

[0009] This invention provides a method for coordinated scheduling of computing power and electricity, comprising the following steps.

[0010] Acquire historical electricity consumption data of computing power hubs in the computing network and historical power generation data of power supply points in the computing network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types; The historical electricity consumption data is input into a pre-trained AAT model to obtain the electricity consumption prediction results for the computing power hub output by the pre-trained AAT model. The historical power generation data is input into a pre-trained STGCN network to obtain the power generation prediction results for the power supply point output by the pre-trained STGCN network. The power consumption forecast results of the computing power hub and the power generation forecast results of the power supply points are input into the pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the corresponding power allocation of each power supply point of various energy types.

[0011] The present invention also provides a computing power and power coordinated scheduling device, comprising the following modules: The data acquisition module is used to acquire historical electricity consumption data of computing power hubs in the computing network and historical power generation data of power supply points in the computing network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types; The computing power prediction module is used to input the historical electricity consumption data into a pre-trained AAT model to obtain the electricity consumption prediction results for the computing power hub output by the pre-trained AAT model. The power prediction module is used to input the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction results for the power supply point output by the pre-trained STGCN network. The computing power collaborative scheduling module is used to input the power consumption forecast results of the computing power hub and the power generation forecast results of the power supply points into a pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the power allocation corresponding to each of the various energy types of power supply points.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the computing power and power coordinated scheduling method as described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the computing power and power coordinated scheduling method as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the computing power and power coordinated scheduling method as described above.

[0015] The present invention provides a method, apparatus, equipment, medium, and product for collaborative scheduling of computing power and electricity. This involves acquiring historical electricity consumption data of computing power hubs and historical power generation data of power supply points in a computing power network. The historical power generation data includes historical power generation data from power supply points of various energy types. The historical electricity consumption data is input into a pre-trained AAT model to obtain the electricity consumption prediction results for the computing power hubs output by the pre-trained AAT model. The historical power generation data is input into a pre-trained STGCN network to obtain the power generation prediction results for the power supply points output by the pre-trained STGCN network. The electricity consumption prediction results for the computing power hubs and the power generation prediction results for the power supply points are input into a pre-trained computing power scheme prediction model to obtain a computing power allocation scheme output by the pre-trained computing power scheme prediction model. The computing power allocation scheme includes the electricity demand of the computing power hubs and the corresponding power allocation amounts for each of the various energy types of power supply points. This method utilizes the AAT model to mine long-term dependencies and complex patterns in computing power data, gaining insights into the computing power usage patterns in different regions. Based on the Spatiotemporal Graph Convolutional Network (STGCN) algorithm, it comprehensively and accurately predicts the power generation patterns of different regions and types of electricity, providing effective data support for computing-power collaborative scheduling. In terms of computing-power collaborative scheduling algorithm design, based on a pre-trained computing-power scheme prediction model, it generates accurate computing-power collaborative scheduling schemes, enabling precise prediction of computing power demand. This allows enterprises to plan resources in advance and flexibly adjust the number of server clusters in the computing power hub based on prediction results, avoiding a large number of devices idling during periods of low computing power demand, thereby effectively reducing power consumption and equipment wear and tear, and significantly reducing operating costs. Simultaneously, it accurately controls power generation and consumption patterns, optimizing power generation plans. In terms of new energy, accurate prediction can improve the absorption capacity of new energy sources such as wind and solar power, reduce energy waste, and improve the overall resource utilization efficiency of the computing network. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the computing power and power coordinated scheduling method provided by the present invention.

[0018] Figure 2 This is a network structure diagram of the STGCN network provided by the present invention.

[0019] Figure 3 This is a schematic diagram of the training process of the pre-trained GAN network provided by the present invention.

[0020] Figure 4 This is a schematic diagram of the network structure of the improved GAN network model provided by the present invention.

[0021] Figure 5 This is a schematic diagram of the computing power and power coordinated scheduling device provided by the present invention.

[0022] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] The following is combined Figures 1-6 Specific embodiments of the present invention are described below.

[0025] Figure 1 This is a flowchart illustrating the computing power and power coordinated scheduling method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following steps.

[0026] Step 101: Obtain historical electricity consumption data of the computing power hub in the computing network and historical power generation data of the power supply points in the computing network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types.

[0027] It should be noted that this application is primarily used to predict the electricity demand of computing hubs and the electricity supply of power plants of various energy types within a specific region over a future period. This region includes a computing network comprised of computing hubs and power supply points; specifically, it predicts the computing power usage patterns of the computing hubs, the electricity usage patterns of the computing hubs, and the power generation patterns of the power supply points. The supply and use of electricity involve multiple energy types, including thermal power, hydropower, wind power, solar power, nuclear power, and biomass power.

[0028] Therefore, it is first necessary to comprehensively collect historical computing power usage data of computing power hubs in a certain region (e.g., region A) and historical power generation data of power supply points corresponding to various energy types.

[0029] This includes historical computing power usage data, such as the computing power hub's workload, computing power resource utilization rate, and task type distribution over different time periods in the past. It also delves into detailed fluctuation data of computing power demand during specific industrial projects or internet activities. Based on this historical computing power usage data, historical electricity consumption data corresponding to various energy types used by the computing power hub can be obtained.

[0030] Historical power generation data for power supply points includes historical power generation data for power supply points of various energy types (thermal power, wind power, hydropower, solar power, etc.), mainly including: power generation, natural environmental data affecting power generation (e.g., environmental data such as light intensity and sunshine duration affecting solar power generation, and meteorological data such as wind speed and wind direction affecting wind power generation), operating data of power generation equipment, and power cost fluctuation data.

[0031] In addition, it is also necessary to obtain information on the topology of the power grid and computing network in the region.

[0032] The historical electricity consumption data and historical power generation data mentioned above were preprocessed, including data cleaning, data standardization, and data normalization. Detailed steps are described in the training process section below.

[0033] Step 102: Input the historical electricity consumption data into the pre-trained AAT model to obtain the electricity consumption prediction results for the computing power hub output by the pre-trained AAT model.

[0034] Among them, the AAT (Adaptive Attention Transformer) model is a Transformer architecture that incorporates an adaptive attention mechanism. Building upon the traditional Transformer, it further enhances the flexibility and adaptability of the attention mechanism. This model can dynamically adjust attention weights based on the features of the input data, thereby better capturing key information in the data and improving the model's performance across various tasks.

[0035] The AAT model consists of an input layer, an adaptive attention layer, a feedforward neural network layer, a normalization layer, a residual connection layer, and an output layer.

[0036] Input layer: The preprocessed historical electricity consumption data described above is used as input. The input dimension is determined by the number of features in the historical electricity consumption data (i.e., computing power data), such as task volume, utilization rate, task type encoding, and other features. At the same time, time-series computing power data is processed using time encoding to help the model capture the time dependencies of the data.

[0037] Adaptive attention layer: By calculating the attention score of data at different locations, the importance of each location in the model processing is determined, thereby adaptively adjusting the attention weights.

[0038] The process begins by calculating the query, key, and value matrices. Then, a dot product is used to obtain the corresponding scores, which are then normalized using the softmax function to arrive at the attention weight for each position. The attention weight for each position is then calculated. The formula for calculating it is:

[0039] (1) Where Q is the query matrix, K is the key matrix, and V is the value matrix. denoted as the dimension of the key matrix.

[0040] When processing computing power data features related to specific industrial projects or internet activities, giving them higher attention weights can make the model more focused on key information and uncover more accurate patterns in computing power usage.

[0041] Finally, the features output by the adaptive attention layer are subjected to nonlinear transformation and feature extraction through multiple fully connected layers, i.e., feedforward neural network layers, to learn more complex patterns and finally output the corresponding results.

[0042] The pre-trained AAT model described above was obtained through the following steps: (1) Obtain the power consumption sample data of the computing power hub in the computing power network (the power consumption sample data is also the computing power usage sample data), and the computing power sample label corresponding to the power consumption sample data (i.e. the actual computing power usage data at a certain point in time). (2) Data preprocessing of the above-mentioned computing power usage sample data (electricity consumption sample data) and computing power sample labels: including, using data integrity index I to evaluate data quality: ; (2) The data integrity index I must reach a preset value (e.g., 0.95) or higher. The collected data must then undergo rigorous cleaning. The 3σ principle can be used to focus on data points deviating from the mean by more than three standard deviations. If anomalies are identified, corrections should be made through verification with relevant departments and reference to similar historical data. Simultaneously, records with a large number of missing values ​​should be deleted if they cannot be filled in a reasonable way to ensure data quality. In addition to checking the accuracy and reasonableness of the data, statistical methods should be used to identify outliers. Normalization methods should be used to transform the computational data into a uniform scale, applying the following normalization formula: (3) Mapped to the interval [0, 1], where It uses sample data for raw computing power. and These are the minimum and maximum values ​​in the dataset, respectively.

[0043] For certain computing power usage sample data (electricity consumption sample data) with seasonal or periodic characteristics, a seasonal decomposition method can also be used, utilizing a time series decomposition model: (4) The data is decomposed into trend, seasonality, and residual components, among which... It is the time series corresponding to the sample data used by the original computing power. It is a trend item. It is a seasonal item. These are residual terms. We standardize these three terms separately to better represent the inherent characteristics of the data and avoid model training instability caused by excessive differences in data volume.

[0044] (3) Finally, a stratified sampling method is used to divide the preprocessed data into training set, validation set and test set to ensure that the data distribution of each subset is similar to the overall data distribution, reduce model training bias and improve the generalization ability of the model.

[0045] (4) Construct the AAT model, input the preprocessed computing power usage sample data (electricity consumption sample data) into the AAT model, and adjust the number of neurons and layers of the fully connected layer to adapt to the learning needs of computing power data patterns with different complexities. The residual connection method is adopted to avoid the gradient vanishing problem, improve the training efficiency and stability of the model, and finally output the corresponding electricity consumption prediction results.

[0046] (5) This embodiment also uses a transfer learning training mode: a Transformer model pre-trained in a relevant domain is found, and its parameters are transferred to the constructed AAT model, mainly including the weights of the attention layer and the fully connected layer. During the transfer process, the transferred parameters are fine-tuned according to the distribution characteristics of the computing power task type, so that the pre-trained model can better adapt to the computing power data. The untransferred AAT parameters are randomly initialized using the Xavier method, and the initialization range is automatically adjusted according to the input and output dimensions of the neurons, so that the parameters can be better distributed in the early stage of training, thus accelerating the convergence speed of the model.

[0047] (6) Based on the nature of the computing power prediction task, the MSE (Mean Squared Error) loss function is selected to measure the difference between the model's predicted value and the true value. The formula for the MSE loss function is: (5) in, It is the actual value. Here, n is the predicted value, and n is the number of samples. Regularization terms, such as L1 or L2 regularization, should also be introduced to prevent overfitting and to constrain the model's complexity, making the model more robust when learning patterns in computing power data.

[0048] (7) Dynamically Adjusting the Learning Rate: Considering the strong volatility and seasonality of the computing power usage sample data (electricity consumption sample data), the Adagrad-Delta optimizer with a dynamic adjustment factor was introduced during training to automatically adjust the learning rate and update the model parameters. When the computing power data (electricity consumption sample data) fluctuates significantly, the learning rate adjustment is dynamically increased by monitoring the second moment of the gradient, enabling the model to adapt to data changes more quickly; when the data fluctuation is small, the learning rate adjustment is decreased to ensure the stability of model training. During training, the model's predicted values ​​are calculated through forward propagation, and then the predicted values ​​are calculated based on the loss function. Calculate the loss value. Next, use the backpropagation algorithm to calculate the gradient of the loss value with respect to the model parameters, and use the Adagrad-Delta optimizer to update the model parameters based on the gradient. It is also necessary to periodically evaluate the model performance, observe the loss value and other evaluation metrics, and take measures such as adjusting the model structure, reducing the number of training epochs, or using regularization to prevent overfitting and maintain the model's generalization ability.

[0049] (8) AAT model evaluation: Define a multi-index evaluation system and a scenario-based evaluation system. Introduce MAPE (Mean Absolute Percentage Error) to reflect the relative error between the predicted value and the actual value. The formula for calculating MAPE is:

[0050] ; (6) in, It is the actual value. Here, n represents the predicted value, and n is the number of samples. Additionally, peak error is introduced to focus on the prediction accuracy of the peak portion of the computing power data. Furthermore, the model is evaluated based on different application scenarios of the computing power data, and its performance in different scenarios is analyzed.

[0051] (9) AAT model optimization: Optimization methods include.

[0052] (9.1) The model parameters can be adjusted according to the scenario to make them conform to the characteristics of the scenario.

[0053] (9.2) Alternatively, the AAT model can be weighted and fused with time series-based ARIMA (Autoregressive Integrated Moving Average) models and GARCH (Generalized Autoregressive Conditional Heteroskedasticity) models for fluctuating data, so as to make full use of the advantages of each model.

[0054] (9.3) Alternatively, a reinforcement learning mechanism can be introduced to optimize the AAT model. The prediction process of the model can be regarded as a decision-making process of an agent in the environment. The environmental state is the feature of the computing power data, the action is the prediction result of the model, and the reward is determined according to the difference between the prediction result and the true value. By continuously training the agent, it can make better prediction decisions based on the characteristics of the computing power data.

[0055] Step 103: Input the historical power generation data into the pre-trained STGCN network to obtain the power generation prediction results of the power supply point output by the pre-trained STGCN network.

[0056] STGCN (Spatial-Temporal Graph Convolutional Network) is a deep learning model used to process graph-structured data with spatiotemporal dependencies. STGCN combines spatial graph convolution and temporal convolution to capture the spatial and temporal features of the data. The structure of the STGCN network is as follows: Figure 2 As shown, the STGCN network mainly consists of two spatiotemporal convolutional modules and an output layer ( Figure 2 As shown in the left column of the diagram, the structure of the spatiotemporal convolutional module (ST-Conv Block) is as follows: Figure 2 As shown in the middle column, each spatiotemporal convolutional module includes two temporal convolutional modules (Temporal Gated Conv) and one spatial graph convolutional module (Spatial Graph-Conv).

[0057] In another embodiment, step 103 specifically includes: constructing a spatiotemporal graph structure based on the power grid topology; wherein, nodes in the spatiotemporal graph structure represent spatial entities in the power grid topology, and edges in the spatiotemporal graph structure represent spatial relationships between the spatial entities; each node carries time-series relationship data obtained from the historical power generation data; inputting the spatiotemporal graph structure into the spatiotemporal graph convolutional layer of the STGCN network to obtain time-series features output by the spatiotemporal graph convolutional layer; inputting the time-series features into a multi-model fusion layer corresponding to the energy type according to the energy type to obtain fused features output by the multi-model fusion layer; wherein, the multi-model fusion layer contains at least two neural network models, each of which is used to extract features from the time-series features and output a first feature; the multi-model fusion layer is used to perform weighted fusion on each of the first features to obtain the fused feature; and inputting the fused feature into a fully connected layer to obtain the power generation prediction result output by the fully connected layer.

[0058] Specifically, the historical power generation data of the power supply points of various energy types obtained above are input into the pre-trained STGCN network to obtain the power generation prediction results output by the pre-trained STGCN network.

[0059] The pre-trained STGCN network described above was obtained through the following steps.

[0060] (1) Data preprocessing: Obtain historical power generation data samples (including natural environmental data affecting power generation, power generation equipment operation data, power cost fluctuation data, etc.) and their corresponding sample labels (i.e., actual power generation) from each power supply point. Use the data integrity index I (as shown in Formula 2 above) to evaluate the data quality of the historical power generation data samples, requiring I to reach above 0.95; then use the K-nearest neighbor algorithm to process missing numbers. For any missing value In its k nearest neighbor data points In the process, the distance weight is calculated. .

[0061] ; (7) in, For distance measurement function, Let be the i-th nearest neighbor data point; the distance metric function can use Euclidean distance; the missing values ​​are estimated by weighted averaging of the distance weights, i.e.

[0062] ; (8) Use the 3σ principle to correct the data and comprehensively check its integrity and accuracy.

[0063] (2) Data Feature Extraction and Encoding: Extract the spatiotemporal features of historical power generation data samples, including periodic time features and time point features, to reflect the changing trend of historical power generation data samples over time. Then, construct a graph structure based on the power grid topology, using power plants, substations, and power consumption areas as nodes of the graph, and transmission lines between them as edges, integrating the spatial relationship of power data into the model, fully considering distance factors, and adjusting the connection weights between nodes such as power plants to better capture the spatial correlation of data. For different energy types, feature encoding is performed. For example, for wind power generation, wind speed, wind direction, and other data are encoded, and wind direction is converted into a unique thermal encoding form; for thermal power and hydropower, relevant data such as fuel supply and water level are extracted as features; for solar power generation, data such as light intensity and sunshine duration are normalized and used as features, while introducing health status indicators of power generation equipment as features, such as equipment failure frequency and maintenance cycle. Z-score standardization is used to standardize all feature data, and stratified sampling is used to divide the dataset. The formula for Z-score standardization is:

[0064] (9) Where x is the original value of the historical power generation data sample. Here, u is the standardized value, representing the sample mean of historical power generation data. It is the sample standard deviation. Z-score standardization can avoid training instability caused by differences in the amount of data.

[0065] (3) Constructing an improved STGCN network: The processed spatiotemporal feature data and energy type data are input into the spatiotemporal graph convolutional layer. The spatial structure of the historical power generation data samples is modeled using graph convolution operations, and time series features are captured using temporal convolution operations. A dynamic weight mechanism is also introduced, that is, the mapping relationship between spatiotemporal characteristics and weights is learned through an additional neural network, so that the model can better adapt to the characteristics of power generation data in different regions. A multi-scale graph convolution strategy is adopted, and local and global spatial features of historical power generation data samples are extracted by combining graph convolution kernels of different sizes. The features are output to the multi-model fusion layer, which contains at least two neural network models. When constructing the multi-model fusion layer, an appropriate neural network model is selected to process different types of power generation data. For example, for solar power generation, a hybrid model of CNN and LSTM is used. CNNs are used to extract local features from data such as light intensity, while LSTMs are used to process time-series data and capture long-term dependencies in solar power generation. For wind power, a clustering algorithm based on a variational autoencoder (VAE) is employed. VAEs can encode and decode wind power data, and the clustering algorithm simultaneously uncovers power generation patterns under different wind zones and conditions. A multi-model fusion layer weights and fuses the outputs of these models, and finally, a fully connected layer performs nonlinear transformations and feature extraction on the fused features to learn more complex power data patterns. Ultimately, a power generation prediction result is output based on the prediction task.

[0066] (4) Define the loss function: Initialize the parameters of the STGCN improved model using Xavier initialization, and select an appropriate loss function according to the nature of the prediction task. For regression tasks, such as power generation prediction, select the MSE loss function (as shown in Formula 5 above) or the MAE (Mean Absolute Error) function. );in It is the actual value. Here, n is the predicted value, and n is the number of samples. For classification tasks, such as power fault type classification, the cross-entropy loss function is used. A penalty term also needs to be introduced to address the characteristics of power data, imposing additional penalties on samples with large prediction errors to improve the model's prediction accuracy.

[0067] (5) STGCN Model Training: The Adam optimizer is used to update model parameters. The first and second moment estimates of the gradient are calculated to dynamically adjust the learning rate, minimize the loss function, and flexibly adjust the learning rate according to the number of training epochs and the performance on the validation set. During training, adjusting the model's hyperparameters is also crucial, including the learning rate, the size of the graph convolution kernel, and the number of neurons in the LSTM hidden layer. Grid search, random search, or Bayesian optimization methods are used to find the optimal hyperparameter combination to train the model. The performance changes of the model on the validation set are observed, and the optimal hyperparameter combination is selected for final model training. Model performance is evaluated regularly, and the loss value and other evaluation metrics are observed. Measures are taken to prevent overfitting and maintain the model's generalization ability.

[0068] (6) STGCN model evaluation: The trained STGCN model is evaluated using test set data. The accuracy, recall, mean squared error, mean absolute error, mean absolute percentage error (MAPE) and other indicators of the model are calculated to comprehensively evaluate the model’s predictive performance on the power generation and consumption patterns of different regions and different types of energy. In addition, the model’s stability and adaptability are comprehensively evaluated in combination with different scenarios, such as different seasons and different peak electricity consumption periods.

[0069] (7) STGCN Model Optimization: Based on the evaluation results, the model is further optimized. Model parameters can be adjusted to improve the prediction performance in certain regions or for certain power generation types, such as increasing the weight of poorly performing regions in graph convolution, or adjusting the feature encoding methods related to those regions. For inaccurate predictions of certain power generation types, the corresponding sub-model parameters or structures can be optimized. Collecting more data or using ensemble learning methods can also be considered to improve the model's stability and accuracy.

[0070] Step 104: Input the power consumption prediction results of the computing power hub and the power generation prediction results of the power supply points into the pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the corresponding power allocation of each power supply point of various energy types.

[0071] The pre-trained prediction model for the computing scheme was built beforehand based on a GAN (Generative Adversarial Network) and a reinforcement learning model. A GAN consists of two neural networks: a generator and a discriminator. These two networks compete against each other during training. The generator aims to generate data (such as images and audio) that is as realistic as possible to deceive the discriminator, making it unable to distinguish between generated and real data. The discriminator's task is to determine as accurately as possible whether the input data is real or generated by the generator. During training, the generator continuously learns how to generate more realistic data, while the discriminator continuously learns how to distinguish more accurately. They reinforce each other, ultimately enabling the generator to produce high-quality, realistic data.

[0072] Specifically, in the method of coordinating computing power and power scheduling, a coordinating scheduling algorithm is designed using generative adversarial networks (GANs). This algorithm focuses on mining and utilizing the intrinsic characteristics of computing and power data. Through the GAN architecture, the computing and power data generated by the previous prediction model and historical data are regenerated to generate a large amount of diverse data that conforms to the laws of computing and power coordination.

[0073] Specifically, in real-time computing power collaborative scheduling, GAN networks are used to simulate and predict the current computing power state. The generator can generate potential future state data based on current power supply and computing power demand, combined with historical data and patterns learned by the model. For example, the generator can predict changes in power generation at different power plants, the growth in computing power demand at different computing hubs, and load changes on transmission lines over a future period (e.g., 1 hour, 3 hours). The discriminator then evaluates this generated future state data, calculating the probability of the discriminator classifying the generated data to select the most likely state. These predictions are then input into the reinforcement learning model to assist in making more accurate scheduling decisions. For example, if it is predicted that the computing power demand in a certain area will increase significantly within the next 3 hours, the reinforcement learning model can adjust the power allocation strategy in advance based on the prediction, prioritizing power allocation from nearby power plants or areas with sufficient power supply to ensure adequate power supply.

[0074] The above embodiments involve acquiring historical electricity consumption data of computing power hubs and historical power generation data of power supply points in the computing power network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types; inputting the historical electricity consumption data into a pre-trained AAT model to obtain the electricity consumption prediction results for computing power hubs output by the pre-trained AAT model; inputting the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction results for power supply points output by the pre-trained STGCN network; inputting the electricity consumption prediction results of computing power hubs and the power generation prediction results of power supply points into a pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the electricity demand of computing power hubs and the corresponding power allocation amounts for each of the various energy types of power supply points. This method utilizes the AAT model to mine long-term dependencies and complex patterns in computing power data, gaining insights into the computing power usage patterns in different regions. Based on the Spatiotemporal Graph Convolutional Network (STGCN) algorithm, it comprehensively and accurately predicts the power generation patterns of different regions and types of electricity, providing effective data support for computing-power collaborative scheduling. In terms of computing-power collaborative scheduling algorithm design, based on a computing-power scheme prediction model, it generates accurate computing-power collaborative scheduling schemes that can accurately predict computing power demand, enabling enterprises to plan resources in advance. Based on the prediction results, the number of server clusters in the computing power hub can be flexibly adjusted to avoid a large number of devices running idle during periods of low computing power demand, thereby effectively reducing power consumption and equipment wear and tear, and significantly reducing operating costs. Simultaneously, it accurately controls power generation and consumption patterns, optimizing power generation plans. In terms of new energy, accurate prediction can improve the absorption capacity of new energy sources such as wind and solar power, reduce energy waste, and improve the overall resource utilization efficiency of the computing network.

[0075] In one embodiment, such as Figure 3 As shown, Figure 3 A schematic diagram illustrating the training process of the prediction model for the power supply scheme is shown. The steps preceding step 101 above include...

[0076] Step 301: Obtain the electricity consumption prediction sample for the computing power hub output by the pre-trained AAT model, and obtain the actual electricity consumption value corresponding to the electricity consumption prediction sample.

[0077] Specifically, this step can be derived from the detailed description of step 102 above, and will not be repeated here.

[0078] Step 302: Obtain the power generation prediction sample for the power supply point output by the pre-trained STGCN network, and obtain the actual power generation value corresponding to the power generation prediction sample.

[0079] Specifically, this step can be derived from the detailed description of step 103 above, and will not be repeated here.

[0080] Step 303: Construct an electricity consumption prediction sample matrix based on the electricity consumption prediction sample, and construct a power generation prediction sample matrix based on the power generation prediction sample; (1) Constructing time series matrices (including electricity consumption prediction sample matrix and power generation prediction sample matrix): Arrange the electricity consumption prediction samples of each computing power hub output during the training process corresponding to step 102 above, and the power generation prediction samples of each power supply point output during the training process corresponding to step 103 above, in chronological order, and construct time series matrices respectively. Assume that the time series corresponding to the power generation prediction sample is... The time series corresponding to the electricity consumption prediction samples is , where m and n are the number of time points for the power generation prediction samples at each power supply point and the number of time points for the power consumption prediction samples at the computing power hub, respectively. To better correlate the power generation patterns of each power supply point, a multi-dimensional time series matrix is ​​constructed for different energy types and regions in the power generation prediction samples. For example, for l power supply points and k energy types, the power generation prediction sample matrix for the power supply points can be represented as follows.

[0081] (10) in, This represents the power generation data at the i-th time point, the j-th power supply point, and the k-th power generation type. m indicates that there are m time points, l indicates that there are l power supply points, and K indicates that there are K energy types.

[0082] Step 304: Use the DTW algorithm to align and associate the electricity consumption prediction sample matrix and the power generation prediction sample matrix in chronological order to obtain the optimal time warp path; the optimal time warp path represents the temporal correspondence of the matrix elements corresponding to the minimum cumulative distance between the electricity consumption prediction sample matrix and the power generation prediction sample matrix.

[0083] The Dynamic Time Warping (DTW) algorithm is used to measure the similarity between two time series. It effectively addresses phase differences or time distortions in time series along the time axis. The DTW algorithm finds the optimal alignment path between two time series, maximizing their cumulative distance (or similarity). This path is an index mapping that specifies which points in the two series correspond on the time axis.

[0084] Specifically, after constructing the computing power and power prediction models (the computing power prediction model is used to predict the power consumption forecast of the computing power hub, and the power prediction model is used to predict the power generation forecast of the power supply point), a computing power data association system needs to be constructed. This embodiment uses the DTW algorithm to associate computing power data, which mainly consists of the following steps.

[0085] Calculate the distance matrix: Define a distance metric function to calculate the distance between the power generation prediction sample matrix of the power supply point and the power consumption prediction sample matrix of the computing hub at each time point. In addition to the commonly used Euclidean distance (the Euclidean distance for two time points i and j...), In addition, considering the multidimensionality and complexity of the data, Mahalanobis distance is used. Let the covariance matrix of the power data (i.e., the power generation prediction sample matrix of power supply points) be... For power data points and computing power data points Mahalanobis distance By traversing all time points of the power data (i.e., the power generation prediction sample matrix of power supply points) and the computing power data (i.e., the power consumption prediction sample matrix of computing power hubs), a... The distance matrix D, where This allows for better consideration of data correlation and scale differences, improving the accuracy of distance calculations.

[0086] Dynamic Programming Solution: The optimal time warp path is found on the distance matrix D using dynamic programming (i.e., the DTW algorithm). The solution method is as follows: Define a cumulative distance matrix M, whose size is also... .initialization The following formula is obtained.

[0087] ; (11) To more accurately correlate the power generation patterns of each power supply point, weight parameters are introduced during the dynamic programming process. and Different weights are assigned to different power supply points and energy types. For example, regions with a high proportion of renewable energy generation are given a higher weight when calculating cumulative distance. (12) in, and Based on the importance of power supply points and energy types, methods such as the Analytic Hierarchy Process (AHP) are used to determine them. This is achieved by backtracking through the cumulative distance matrix. ,from We begin by determining the optimal time warp path (meaning the path where the cumulative distance between two sequences is minimized) based on the path with the minimum cumulative distance. This path represents the best alignment of power data (i.e., the generation prediction sample matrix of power supply points) and computing power data (i.e., the electricity consumption prediction sample matrix of computing power hubs) in the time dimension.

[0088] Step 305: Based on the correlation of elements at each time point in the optimal time warp path, obtain the power consumption prediction sample correlation index; the power consumption prediction sample correlation index represents the temporal coupling strength between the power consumption prediction sample and the power generation prediction sample. Among them, the correlation indicators for power prediction samples can be some statistical indicators, such as correlation coefficient, Pearson correlation coefficient, etc.

[0089] Specifically, the coupling relationship is determined: based on the found optimal time warp path, the correspondence between power generation data and power consumption data at specific points in time is analyzed. For different energy types and different power supply points in the power generation data, correlation analysis is performed with the power consumption data. For example, for regions with a high proportion of solar power generation, if within a certain time period in the optimal time warp path, the increase in solar power generation in this region frequently corresponds to the increase in the amount of AI training tasks in the power consumption data, and the correlation coefficient reaches a significant level through hypothesis testing (such as t-test) (e.g., ... If this is the case, then we can initially determine that there is a strong coupling relationship between the two during this period. By statistically analyzing these correspondences, we can uncover potential coupling patterns between power generation data and power consumption data, while also considering the characteristics of different power supply points and energy types to achieve more precise correlation and coupling.

[0090] Step 306: Based on the correlation between the actual electricity consumption value and the actual power generation value at each time point, obtain the correlation index of the actual electricity consumption value.

[0091] The specific calculation process for this step is similar to step 305 above, and will not be repeated here.

[0092] Optionally, visualization analysis can also be performed: power generation and consumption data can be visualized in aligned chronological order using multi-dimensional visualization methods. For example, a 3D line chart can be used to show the relationship between power generation and computing power tasks at different power supply points over time, with colors used to distinguish different energy types; or a parallel coordinate graph can be used to display multiple dimensions of power generation data (such as power generation, power supply points, and power generation type) and multiple dimensions of consumption data (such as computing power tasks, task types, and computing power nodes) on the same graph, intuitively showing the changing trends and corresponding relationships of the two over time, further verifying and analyzing potential coupling relationships. Through visualization, some abnormal correspondences can also be discovered, such as a correlation between power generation and computing power demand in a certain power generation area that does not conform to the overall trend within a specific time period, so as to further check the accuracy of the data or algorithm.

[0093] Correlation Calculation: Calculate the correlation coefficient between the aligned power generation and consumption data. Besides the Pearson correlation coefficient, the Spearman rank correlation coefficient can also be used. The formula for calculating the Pearson correlation coefficient r is...

[0094] ; (13) The Pearson correlation coefficient r is used to measure the degree of linear correlation between two variables, and its value ranges from [-1, 1].

[0095] The formula for calculating the Spearman rank correlation coefficient is as follows.

[0096] (14) in, The correlation coefficient is the rank difference between two variables x and y, used to measure the degree of monotonic correlation between the two variables. By calculating various correlation coefficients, the coupling strength between the two can be assessed more comprehensively. If the correlation coefficient is close to 1 or -1, it indicates a strong linear or monotonic correlation between the two; if it is close to 0, it indicates a weak correlation, but other non-linear coupling relationships may exist.

[0097] Model Validation: To verify the reliability of the coupling relationships obtained through the DTW (Dynamic Time Warping) algorithm, leave-one-out cross-validation (LOOCV) is employed. The data is divided into N subsets. Each time, N-1 subsets are selected as the training set, and the remaining subset is used as the test set. The DTW algorithm is run on the training set to obtain the coupling relationships, and then these relationships are verified on the test set to ensure they still hold. The stability and accuracy of the model are evaluated through N iterations. If the validation results are unsatisfactory, the data preprocessing steps, the choice of distance metric function, or the implementation process of dynamic programming can be re-examined, and corresponding adjustments and optimizations can be made. For example, if it is found that the data from a certain power generation area (or power supply point) has a significant impact on the coupling relationships and the validation effect is poor, the preprocessing method for the data in that area can be re-examined, or the weight parameters for that area in distance calculation and dynamic programming can be adjusted.

[0098] Application of Correlated Data: The coupling relationship between power generation and consumption data obtained through DTW (Dynamic Time Warping) algorithm analysis is applied to subsequent collaborative scheduling. Based on the previously constructed prediction model, the potential changes in power generation demand when computing power demand (i.e., consumption data) increases within a certain time period are predicted, thereby adjusting the power supply plan in advance and achieving more rational computing-electricity collaborative scheduling. The coupling relationship is used as an important constraint or reference factor when designing the collaborative scheduling algorithm. For example, when allocating power resources, the coupling relationship matrix is ​​used... (in (This represents the coupling strength between the i-th computing node and the j-th power supply point). An optimization algorithm is used to solve the power resource allocation problem, prioritizing the needs of regions or tasks that are strongly coupled with computing power requirements, so as to improve the utilization efficiency of power resources and the stability of computing power services.

[0099] Continuous optimization and updates: As time progresses and data accumulates, the DTW algorithm is periodically rerun to update the coupling relationship between power generation and consumption data. This is because the power system and computing system may be affected by various factors, such as the integration of new energy power generation and the emergence of new computing task types, leading to changes in their coupling relationship. Monitoring indicators for changes in the coupling relationship are established, such as the coupling stability index S.

[0100] ; (15) in, and These are the coupling coefficients before and after the update, respectively. When S exceeds a certain threshold (e.g., 0.1), the DTW algorithm is triggered to run again and the coupling coefficients are updated. Through continuous optimization and updates, it can be ensured that the data support provided for collaborative scheduling is always targeted and scientific.

[0101] Step 307: Based on the correlation index of the power prediction sample and the correlation index of the actual power value, construct the joint loss function of power calculation.

[0102] The joint loss function of computing and power can be the computing and power consistency loss. .

[0103] Computer consistency loss This is used to ensure that the generated power supply prediction samples are consistent with the real data (i.e., the actual power supply values) in terms of their collaborative relationships. For example, for the matching relationship between power supply (generation data) and computing power demand (consumption data), the loss function is constructed as follows by calculating the differences in power-computing power correlation indicators between the generated data (i.e., power consumption prediction samples and generation prediction samples) and the real data (including actual power consumption values ​​and actual generation values).

[0104] ; (16) in, and These are the correlation indicators of the actual power calculation value and the correlation indicators of the power calculation prediction sample.

[0105] Step 308: Based on the joint loss function of computing and power, train the computing and power scheme prediction model to obtain the pre-trained computing and power scheme prediction model; wherein, the computing and power scheme prediction model is constructed based on a GAN network.

[0106] Specifically, using the aforementioned computational consistency loss The GAN network is trained to obtain the pre-trained GAN network. For example, the optimal time warp path is input into the generator of the GAN network to obtain the data generated by the generator. Then, the data generated by the generator, along with the actual power generation and power consumption values, are input into the discriminator of the GAN network to obtain the discriminant output. The discriminant outputs the discrimination result, and the consistency loss of power generation is calculated using the correlation index of the actual power generation value and the correlation index of the power generation prediction sample. Adjust the GAN network parameters until the aforementioned computational consistency loss is reached. If the preset conditions are met, a trained generator is obtained. The trained generator is then combined with a trained reinforcement learning model to form the prediction model for the aforementioned power calculation scheme.

[0107] The above embodiments utilize the coupling relationship between power data and computing power data obtained through DTW algorithm analysis, and apply this relationship to the subsequent training process of the collaborative scheduling model. This allows the coupling relationship to be used as an important constraint or reference factor during the training of the collaborative scheduling model, providing more reliable data support for collaborative scheduling and improving the accuracy of the scheduling scheme.

[0108] In another embodiment, step 307 includes: constructing a computational consistency loss function based on the computational prediction sample association index and the computational real value association index; constructing a data diversity loss function based on the differences between the data in the dataset generated by the generator in the GAN network; and constructing the computational joint loss function according to the computational consistency loss function, the data diversity loss function, and the adversarial loss function of the GAN network.

[0109] Specifically, this embodiment describes the training process for optimizing computing and power characteristics, employing a joint loss function that includes adversarial loss, computing-power consistency loss, and data diversity loss, as well as an adaptive training strategy to ensure that the model can accurately learn the computing-power cooperative mode.

[0110] The joint computational-electronic loss function L: In this embodiment, the loss function is redesigned to consider not only the similarity between generated data and real data, but also the relevant constraints of computational-electronic collaboration. The joint computational-electronic loss function L consists of three parts.

[0111] (1) Traditional combat losses , is used to measure the adversarial effect between the generator and the discriminator.

[0112] ; (17) (2) Computational consistency loss function This is used to ensure that the generated computing power data is consistent with the real data in terms of collaborative relationships. For example, for the matching relationship between power supply and computing power demand, this loss is constructed by calculating the difference between the power-computing power correlation indicators in the generated data and the real data.

[0113] ; (18) in, and These are the power calculation correlation indicators for real data and generated data, respectively.

[0114] (3) Data diversity loss function This is used to ensure that the generated data has sufficient diversity. This can be achieved by calculating the differences between the data in the generated dataset.

[0115] (19) The final computational-electrical joint loss function L is:

[0116] ; (20) in, , , These are weighting coefficients, which are adjusted according to the actual situation.

[0117] An adaptive training strategy is employed: during training, training parameters are dynamically adjusted based on the performance of the generator and discriminator. For example, if the discriminator becomes too powerful, making it difficult for the generator to produce effective data, the discriminator's learning rate is reduced. This can be achieved using exponential decay, such as... ,in It is the learning rate at step t. It is the initial learning rate. This is the decay coefficient (typically between 0.9 and 0.99). Alternatively, increase the number of training steps for the generator, allowing it to iterate and update multiple times during each training iteration, while the discriminator updates only once, to balance the training progress of both. Conversely, when the data generated by the generator is too similar and lacks diversity, increase the weight of the data diversity loss. It can be dynamically adjusted based on the diversity metrics of the generated data (such as by calculating the entropy of the generated data). When the diversity index falls below a set threshold, it will be increased by a certain percentage. The value of . Or adjust the generator's network structure, such as increasing the number of convolutional layers or neurons in the generator, to increase the generator's expressive power and enable it to generate more diverse data.

[0118] The above embodiments utilize the joint computation and power loss function to adaptively train the GAN network, which can improve the quality of the generated data and the discriminator's ability to distinguish the data, ultimately enabling the trained GAN network to generate a more reliable computation and power allocation scheme.

[0119] In one embodiment, the above-mentioned power consumption prediction model is obtained based on a pre-trained GAN network and a pre-trained reinforcement learning model; the generator and discriminator of the pre-trained GAN network each include a power consumption feature embedding layer and a multi-scale convolutional module. Step 104 includes: inputting the power consumption prediction result of the computing power hub and the power generation prediction result of the power supply point into the power consumption feature embedding layer in the generator, so that the power consumption feature embedding layer performs feature extraction and feature embedding on the power consumption prediction result and the power generation prediction result to obtain power consumption features; and inputting the power consumption prediction result into the power supply network. The features are input into the multi-scale convolutional module of the generator to obtain the concatenated features output by the multi-scale convolutional module; wherein, the multi-scale convolutional module contains multiple convolutional layers; each convolutional layer uses a different kernel size; the multi-scale convolutional module is used to extract features from the computing power features using each convolutional layer to obtain the second features output by each convolutional layer, and the second features are concatenated to obtain the concatenated features; the concatenated features are input into a pre-trained reinforcement learning model to obtain the computing power allocation scheme output by the pre-trained reinforcement learning model.

[0120] This embodiment employs an improved GAN network model. A schematic diagram of the improved GAN network model's network structure is shown below. Figure 4 As shown.

[0121] Specifically, such as Figure 4 As shown, a computing power feature embedding layer and a multi-scale convolution module are introduced into the generator and discriminator. The computing power feature embedding layer is used to deeply capture the local and global features of power and computing power data. For power data, one-hot encoding is used to convert different energy types (such as thermal power, hydropower, wind power, solar power, etc.) into vector form, and then feature extraction and embedding are performed through a multilayer perceptron (MLP). For power generation, after normalization, feature extraction and embedding are performed through small convolutional kernels of a convolutional neural network (CNN) to capture the local variation features of power generation. For transmission line status data, after standardizing indicators such as line load rate and loss rate, a recurrent neural network (RNN) is used for feature embedding, because RNN can handle transmission line status data with time series characteristics well. For computing power data, computing power demand is first normalized, and then feature embedding is performed through a fully connected layer to reflect the overall scale of computing power demand. For computing power task data, a one-hot encoding method combined with an MLP (Multilayer Perceptron) is used for feature embedding, similar to that used for energy type data. Computing power utilization is standardized, and its local features are extracted using a CNN and then embedded. Through these operations, the model can deeply capture the intrinsic characteristics of computing power data, laying the foundation for generating high-quality data subsequently.

[0122] In addition, multi-scale convolution modules are introduced in the generator and discriminator. These modules consist of convolutional layers with kernels of different sizes (e.g., kernel sizes of 1×1, 3×3, and 5×5). The 1×1 kernel is mainly used to perform linear transformations on the data, the 3×3 kernel can capture medium-scale features of the data, such as the collaborative relationship features between power generation equipment in power data and the changing features of computing power demand in computing power data, and the 5×5 kernel is used to obtain large-scale features of the data, such as the overall power transmission pattern between different regions in the power system and the correlation features between different computing power hubs in the computing power network. The second feature processed by the convolution kernels of different sizes is concatenated to obtain the concatenated feature, which is then processed by batch normalization and activation functions (such as ReLU) before being input into the pre-trained reinforcement learning model to obtain the computing power allocation scheme output by the pre-trained reinforcement learning model.

[0123] The above embodiments, by setting convolutional layers with different kernel sizes, can improve the quality of generated data and the discriminator's ability to distinguish data.

[0124] In one embodiment, after step 104 above, the method further includes: calculating the probability of each power station being selected based on the resources required per unit of electricity for each power station; determining the final power station in the power allocation scheme based on the ranking of the selected probabilities of each power station among all power stations; and / or, when the emergency status index is greater than a preset index threshold, calculating the power allocation amount of the preset key computing power task based on the priority of the preset key computing power task.

[0125] This embodiment improves the scheduling strategy in the power allocation scheme, using a more targeted scheduling strategy based on different actual situations, specifically including...

[0126] 1. Electricity Cost Priority Strategy: Based on real-time cost information from the electricity market (i.e., resources required per unit of electricity), prioritize power plants with lower electricity costs to supply power to the computing hub. Let the electricity price of power plant j be... The power supply is The probability of choosing power station j is... It can be calculated using the softmax function.

[0127] (twenty one) in, This is an adjustment parameter used to control the sensitivity to electricity price differences. During dispatch decisions, power plants are selected based on this probability to reduce electricity procurement costs.

[0128] 2. Emergency Response Strategy: An emergency response strategy will be activated when a sudden power system failure occurs or the computing hub encounters an urgent computing power demand. Let the emergency status indicators be... ,when ( When the threshold is reached, an emergency response is triggered. In this case, priority is given to ensuring the power supply for critical computing tasks, with the priority of critical computing tasks being [missing information]. Electricity distribution It can be calculated using the following formula.

[0129] ; (twenty two) in, N represents the total available power, and N represents the total number of tasks. Simultaneously, by collaborating with other power supply regions or computing hubs, the system can respond to unforeseen circumstances and ensure the stability and reliability of the entire computing system.

[0130] 3. Data-driven precise scheduling strategy: A large amount of computational data is generated using an improved GAN to pre-train a deep reinforcement learning model. In reinforcement learning, the policy network (Actor network) plays a crucial role. Generate action a based on the current state s, with the following parameters: Value Network (Critic Network) It is used to evaluate the value of taking action a in the current state s, and its parameters are: The optimization objective of the policy network is to maximize the cumulative reward, and the parameters are updated using the gradient ascent method. Its gradient is .

[0131] ; (twenty three) in, The performance metric of a strategy is usually expressed as the expected value of cumulative rewards. For policy functions; The gradient of the policy function; [] represents the expected value of state s and action a, where state s is determined according to the policy. State distribution under Sampling, action a according to strategy sampling; It is a strategy The state distribution under [condition]. Parameters of the value network. Update by minimizing the time difference error (TD error), i.e.

[0132] ; (twenty four) in, Let be the loss function of the value network; This is the target value used to update the parameters of the value network; Let be the action value function, representing the state. Take action below The expected cumulative reward that the strategy can obtain; It's an instant reward. It is a discount factor. The next state is... These are possible actions; The next state All possible actions The maximum Q-value, which represents the optimal action value in the next state, is used to calculate the target value. n represents the number of samples. In real-time scheduling, the future state data generated by the GAN... After evaluation and screening by the discriminator, the data is input into the reinforcement learning model to help it make more accurate scheduling decisions.

[0133] 4. Power-Computing Collaborative Optimization Strategy: Let C be the power-computing power matching index, and let the power-computing power matching index in the actual data be... The power-computing power matching index in the generated data is: The consistency loss function is .

[0134] ; (25) Where n is the total number of samples, and when making scheduling decisions, the generated data shows the computing power demand of a certain region. When increasing, assume the current power supply in the area is... According to power supply forecasts Adjusting power distribution The formula is as follows.

[0135] ; (26) Where k is an adjustment coefficient, determined according to actual conditions. This formula adjusts power allocation to ensure that power supply can meet computing power requirements, achieving high efficiency in computing-power collaboration.

[0136] 5. Intelligent Dispatch Strategy: Let the predicted power generation value of the new energy power station j be... The total power generation forecast is The proportion allocated to power supply for computing hubs by new energy power plants. Calculate using the following formula.

[0137] ;(27) in, It is the variance of the predicted value of new energy power generation. It is the average value of the predicted power generation of new energy sources. This is the adjustment coefficient. When the generated data predicts sufficient renewable energy generation, the scheduling strategy is automatically adjusted according to this formula to allocate more renewable energy to the computing hub, maximizing the utilization of green energy while ensuring the stable operation of computing services.

[0138] The above embodiments, by designing different power scheduling schemes for different scenarios, are more accurate and efficient, improve environmental protection, significantly reduce energy consumption in scenarios such as data centers, reduce electricity expenses, and achieve energy conservation and emission reduction goals.

[0139] The computing power and power coordinated scheduling device provided by the present invention is described below. The computing power and power coordinated scheduling device described below and the computing power and power coordinated scheduling method described above can be referred to each other.

[0140] like Figure 5 As shown, Figure 5 A schematic diagram of a computing power and power coordinated scheduling device is provided, which includes the following modules.

[0141] The data acquisition module 501 is used to acquire historical electricity consumption data of computing power hubs in the computing network and historical power generation data of power supply points in the computing network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types; The computing power prediction module 502 is used to input the historical electricity consumption data into the pre-trained AAT model to obtain the electricity consumption prediction result for the computing power hub output by the pre-trained AAT model. The power prediction module 503 is used to input the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction result for the power supply point output by the pre-trained STGCN network. The computing power collaborative scheduling module 504 is used to input the power consumption prediction results of the computing power hub and the power generation prediction results of the power supply points into a pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the power allocation corresponding to each of the various energy types of power supply points.

[0142] In one embodiment, the above-described apparatus further includes a model training unit, used for: Obtain the electricity consumption prediction sample for the computing power hub output by the pre-trained AAT model, and obtain the actual electricity consumption value corresponding to the electricity consumption prediction sample. Obtain the power generation prediction sample for the power supply point output by the pre-trained STGCN network, and obtain the actual power generation value corresponding to the power generation prediction sample; Construct an electricity consumption prediction sample matrix based on the electricity consumption prediction sample; The DTW algorithm is used to align and associate the electricity consumption prediction sample matrix and the power generation prediction sample matrix in chronological order to obtain the optimal time warp path; the optimal time warp path represents the temporal correspondence of the matrix elements corresponding to the minimum cumulative distance between the electricity consumption prediction sample matrix and the power generation prediction sample matrix. Based on the correlation of elements at each time point in the optimal time warp path, a power consumption prediction sample correlation index is obtained; the power consumption prediction index represents the temporal coupling strength between the power consumption prediction data and the power generation prediction results. Based on the correlation between the actual electricity consumption value and the actual power generation value at each time point, the correlation index of the calculated actual electricity value is obtained; Based on the correlation index of the power prediction sample and the correlation index of the actual power value, a joint loss function for power calculation is constructed. Based on the aforementioned joint loss function for computation and power supply, a computation and power supply scheme prediction model is trained to obtain the pre-trained computation and power supply scheme prediction model; wherein, the computation and power supply scheme prediction model is constructed based on a GAN network and a reinforcement learning model.

[0143] In one embodiment, the above-described model training unit is further configured to: Based on the correlation index of the power prediction sample and the correlation index of the actual power value, a power consistency loss function is constructed. Based on the differences between data in the dataset generated by the generator in the GAN network, a data diversity loss function is constructed. The joint computation-power consistency loss function is constructed based on the computation-power consistency loss function, the data diversity loss function, and the adversarial loss function of the GAN network.

[0144] In one embodiment, the power prediction module 503 is further configured to: A spatiotemporal graph structure is constructed based on the power grid topology; wherein, the nodes in the spatiotemporal graph structure are used to represent spatial entities in the power grid topology, and the edges in the spatiotemporal graph structure are used to represent spatial relationships between the spatial entities; each node carries time-series relationship data obtained from the historical power generation data; The spatiotemporal graph structure is input into the spatiotemporal graph convolutional layer of the STGCN network to obtain the time series features output by the spatiotemporal graph convolutional layer; Based on the energy type to which the time series features belong, the time series features are input into a multi-model fusion layer corresponding to the energy type to obtain the fused features output by the multi-model fusion layer; wherein, the multi-model fusion layer contains at least two neural network models, each of which is used to extract features from the time series features and output a first feature; the multi-model fusion layer is used to perform weighted fusion on each of the first features to obtain the fused features; The fused features are input into the fully connected layer to obtain the power prediction result output by the fully connected layer.

[0145] In one embodiment, the computational power scheme prediction model is obtained based on a pre-trained GAN network and a pre-trained reinforcement learning model; the generator of the pre-trained GAN network includes a computational power feature embedding layer and a multi-scale convolution module, and the aforementioned computational power cooperative scheduling module 504 is further used for: The power consumption prediction results of the computing hub and the power generation prediction results of the power supply point are input into the power consumption feature embedding layer in the discriminator, so that the power consumption feature embedding layer performs feature extraction and feature embedding on the power consumption prediction results and the power generation prediction results to obtain power consumption features; The computational power features are input into the multi-scale convolutional module of the discriminator to obtain the concatenated features output by the multi-scale convolutional module; wherein, the multi-scale convolutional module contains multiple convolutional layers; each convolutional layer uses a different kernel size; the multi-scale convolutional module is used to extract features from the computational power features using each convolutional layer to obtain a second feature output by each convolutional layer, and the second features are concatenated to obtain the concatenated features; The splicing features are input into a pre-trained reinforcement learning model to obtain the computing power allocation scheme output by the pre-trained reinforcement learning model.

[0146] In one embodiment, the above-mentioned computing power collaborative scheduling module 504 is further configured to: Based on the resources required per unit of electricity generated by each power station, calculate the probability of each power station being selected; and determine the final power stations in the power allocation scheme based on the ranking of the selected probabilities among all power stations. And / or, When the emergency status indicator exceeds the preset indicator threshold, the power allocation of the preset key computing power task is calculated based on the priority of the preset key computing power task.

[0147] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communications bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other through the communications bus 640. The processor 610 can call logic instructions in the memory 630 to execute a computing power and power coordinated scheduling method. This method includes: acquiring historical power consumption data of computing power hubs and historical power generation data of power supply points in the computing power network; wherein the historical power generation data includes historical power generation data of power supply points of various energy types; inputting the historical power consumption data into a pre-trained AAT model to obtain the power consumption prediction results for the computing power hubs output by the pre-trained AAT model; inputting the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction results for the power supply points output by the pre-trained STGCN network; inputting the power consumption prediction results of the computing power hubs and the power generation prediction results of the power supply points into a pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hubs and the corresponding power allocation amounts for each of the various energy types of power supply points.

[0148] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0149] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the computing power and power coordinated scheduling method provided by the above methods. The method includes: acquiring historical power consumption data of computing power hubs and historical power generation data of power supply points in the computing power network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types; inputting the historical power consumption data into a pre-trained AAT model to obtain the power consumption prediction result for the computing power hub output by the pre-trained AAT model; inputting the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction result for the power supply point output by the pre-trained STGCN network; inputting the power consumption prediction result of the computing power hub and the power generation prediction result of the power supply point into a pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the corresponding power allocation amount for each of the various energy types of power supply points.

[0150] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program is implemented to perform the computing power and power coordinated scheduling method provided by the above methods. The method includes: acquiring historical power consumption data of computing power hubs and historical power generation data of power supply points in the computing power network; wherein the historical power generation data includes historical power generation data of power supply points of various energy types; inputting the historical power consumption data into a pre-trained AAT model to obtain the power consumption prediction result for the computing power hub output by the pre-trained AAT model; inputting the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction result for the power supply point output by the pre-trained STGCN network; inputting the power consumption prediction result of the computing power hub and the power generation prediction result of the power supply point into a pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the corresponding power allocation amount for each of the various energy types of power supply points.

[0151] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0152] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for coordinated scheduling of computing power and electricity, characterized in that, include: Acquire historical electricity consumption data of computing power hubs in the computing network and historical power generation data of power supply points in the computing network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types; The historical electricity consumption data is input into a pre-trained AAT model to obtain the electricity consumption prediction results for the computing power hub output by the pre-trained AAT model. The historical power generation data is input into a pre-trained STGCN network to obtain the power generation prediction results for the power supply point output by the pre-trained STGCN network. The power consumption forecast results of the computing power hub and the power generation forecast results of the power supply points are input into the pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the corresponding power allocation of each power supply point of various energy types.

2. The computing power and power coordinated scheduling method according to claim 1, characterized in that, Before acquiring the historical electricity consumption data of the computing power hub in the computing network and the historical power generation data of the power supply points in the computing network, the following steps are included: Obtain the electricity consumption prediction sample for the computing power hub output by the pre-trained AAT model, and obtain the actual electricity consumption value corresponding to the electricity consumption prediction sample. Obtain the power generation prediction sample for the power supply point output by the pre-trained STGCN network, and obtain the actual power generation value corresponding to the power generation prediction sample; Construct an electricity consumption forecast sample matrix based on the electricity consumption forecast sample, and construct a power generation forecast sample matrix based on the power generation forecast sample; The DTW algorithm is used to align and associate the electricity consumption prediction sample matrix and the power generation prediction sample matrix in chronological order to obtain the optimal time warp path; the optimal time warp path represents the temporal correspondence of the matrix elements corresponding to the minimum cumulative distance between the electricity consumption prediction sample matrix and the power generation prediction sample matrix. Based on the correlation of elements at each time point in the optimal time warp path, the correlation index of the power consumption prediction sample is obtained; the correlation index of the power consumption prediction sample represents the temporal coupling strength between the power consumption prediction sample and the power generation prediction sample. Based on the correlation between the actual electricity consumption value and the actual power generation value at each time point, the correlation index of the calculated actual electricity value is obtained; Based on the correlation index of the power prediction sample and the correlation index of the actual power value, a joint loss function for power calculation is constructed. Based on the aforementioned joint loss function for computation and power generation, a computation and power generation scheme prediction model is trained to obtain the pre-trained computation and power generation scheme prediction model; wherein, the computation and power generation scheme prediction model is constructed based on a GAN network.

3. The computing power and electric power coordinated scheduling method of claim 2, wherein, The step of constructing a joint loss function for power calculation based on the correlation index of the power calculation prediction samples and the correlation index of the actual power calculation values ​​includes: Based on the correlation index of the power prediction sample and the correlation index of the actual power value, a power consistency loss function is constructed. Based on the differences between data in the dataset generated by the generator in the GAN network, a data diversity loss function is constructed. The joint computation-power consistency loss function is constructed based on the computation-power consistency loss function, the data diversity loss function, and the adversarial loss function of the GAN network.

4. The computing power and power coordinated scheduling method according to claim 1, characterized in that, The step of inputting the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction results for the power supply point output by the pre-trained STGCN network includes: A spatiotemporal graph structure is constructed based on the power grid topology; wherein, the nodes in the spatiotemporal graph structure are used to represent spatial entities in the power grid topology, and the edges in the spatiotemporal graph structure are used to represent spatial relationships between the spatial entities; each node carries time-series relationship data obtained from the historical power generation data; The spatiotemporal graph structure is input into the spatiotemporal graph convolutional layer of the STGCN network to obtain the time series features output by the spatiotemporal graph convolutional layer; Based on the energy type to which the time series features belong, the time series features are input into a multi-model fusion layer corresponding to the energy type to obtain the fused features output by the multi-model fusion layer; wherein, the multi-model fusion layer contains at least two neural network models, each of which is used to extract features from the time series features and output a first feature; the multi-model fusion layer is used to perform weighted fusion on each of the first features to obtain the fused features; The fused features are input into the fully connected layer to obtain the power generation prediction result output by the fully connected layer.

5. The computing power and power coordinated scheduling method according to claim 1, characterized in that, The power consumption prediction model is based on a pre-trained GAN network and a pre-trained reinforcement learning model. The generator of the pre-trained GAN network includes a power consumption feature embedding layer and a multi-scale convolutional module. The power consumption prediction results of the computing power hub and the power generation prediction results of the power supply point are input into the pre-trained power consumption prediction model to obtain the power allocation scheme output by the pre-trained power consumption prediction model, including: The power consumption prediction results of the computing hub and the power generation prediction results of the power supply point are input into the power consumption feature embedding layer in the generator, so that the power consumption feature embedding layer performs feature extraction and feature embedding on the power consumption prediction results and the power generation prediction results to obtain power consumption features; The computational power features are input into the multi-scale convolutional module of the generator to obtain the concatenated features output by the multi-scale convolutional module; wherein, the multi-scale convolutional module contains multiple convolutional layers; each convolutional layer uses a different kernel size; the multi-scale convolutional module is used to extract features from the computational power features using each convolutional layer to obtain a second feature output by each convolutional layer, and the second features are concatenated to obtain the concatenated features; The splicing features are input into a pre-trained reinforcement learning model to obtain the computing power allocation scheme output by the pre-trained reinforcement learning model.

6. The computing power and power coordinated scheduling method according to claim 1, characterized in that, The step of inputting the power consumption prediction results of the computing power hub and the power generation prediction results of the power supply point into a pre-trained computing power allocation scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power allocation scheme prediction model includes: Based on the resources required per unit of electricity generated by each power station, calculate the probability of each power station being selected; and determine the final power stations in the power allocation scheme based on the ranking of the selected probabilities among all power stations. And / or, When the emergency status indicator exceeds the preset indicator threshold, the power allocation of the preset key computing power task is calculated based on the priority of the preset key computing power task.

7. A computing power and power coordinated scheduling device, characterized in that, include: The data acquisition module is used to acquire historical electricity consumption data of computing power hubs in the computing network and historical power generation data of power supply points in the computing network; wherein, the historical power generation data includes historical power generation data of power supply points of various energy types; The computing power prediction module is used to input the historical electricity consumption data into a pre-trained AAT model to obtain the electricity consumption prediction results for the computing power hub output by the pre-trained AAT model. The power prediction module is used to input the historical power generation data into a pre-trained STGCN network to obtain the power generation prediction results for the power supply point output by the pre-trained STGCN network. The computing power collaborative scheduling module is used to input the power consumption forecast results of the computing power hub and the power generation forecast results of the power supply points into a pre-trained computing power scheme prediction model to obtain the computing power allocation scheme output by the pre-trained computing power scheme prediction model; the computing power allocation scheme includes the power demand of the computing power hub and the power allocation corresponding to each of the various energy types of power supply points.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the computing power and power coordinated scheduling method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the computing power and power coordinated scheduling method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the computing power and power coordinated scheduling method as described in any one of claims 1 to 6.