Regional power grid operation optimization method and system based on multi-modal diffusion model

Abstract

This invention discloses a method and system for optimizing regional power grid operation based on a multimodal diffusion model. The method includes: constructing a multimodal operation information feature library for the regional power grid; calculating teacher diffusion models and student diffusion models; constructing a hybrid sample feature pool; performing extreme value-aware dynamic iterative knowledge distillation on the student diffusion model using the teacher diffusion model for denoising; and then performing time-step-aware low-bit quantization processing; constructing a calibration dataset; executing a separation quantization strategy on the student diffusion model to generate a quantized diffusion model; constructing a progressive training objective; and fine-tuning the quantized diffusion model with a minimal learning rate in a floating-point computing environment to obtain a lightweight diffusion model, thereby obtaining power grid situation projection results and adjusting the regional power grid operation mode. This invention provides a forward-looking situation awareness means for decision-making on high-risk sensitive lines and guiding differentiated power grid scheduling, thus improving the power grid's disaster prevention and mitigation capabilities.

Description

Technical Field

[0001] This invention mainly relates to the field of power system technology, specifically to a method and system for optimizing the operation of a regional power grid based on a multimodal diffusion model. Background Technology

[0002] With the in-depth construction of new power systems, the high proportion of new energy integration, and the frequent occurrence of extreme weather events (such as typhoons, freezing rain, and extreme heat), the operating environment and topology of the power grid are becoming increasingly complex. To enhance the power grid's security and defense capabilities, it is necessary to integrate multimodal information, such as meteorological data (e.g., ambient temperature and humidity, wind speed, irradiance) and power grid data (e.g., bus voltage, line power flow, switch status), to perform high-precision power grid state identification and situation prediction. In recent years, diffusion models, with their powerful distribution fitting capabilities in multi-step iterative denoising processes, have demonstrated extremely high fidelity and generalization in multimodal data generation and situation prediction tasks, and are gradually becoming the core algorithm for intelligent power grid control.

[0003] However, existing diffusion models face insurmountable technological gaps when deployed to frontline power grid operations (such as substation edge smart gateways, transmission line inspection drones, and distribution area edge computing terminals): First, there is an inherent contradiction between the hardware computing power constraints at the edge and the computing power requirements of large models. The generation process of the diffusion model is essentially a Markov chain, requiring hundreds of iterations of noise estimation, which relies on deep neural networks (such as the U-Net architecture) with an extremely large number of parameters (billions of parameters). In practical engineering, such models heavily depend on the floating-point operation (FP32 / FP16) capabilities of high-performance GPU clusters in the cloud. However, due to limitations in size and power consumption, power grid edge terminals are typically equipped only with NPUs dedicated to low-precision fixed-point operations (such as INT8).

[0004] Secondly, conventional quantization compression methods are severely incompatible with diffusion models. To adapt to NPUs, industry typically employs post-training quantization (PTQ) techniques. However, in the denoising network of a diffusion model, the distribution of output activation values ​​varies significantly across different backward time steps; furthermore, the activation distribution of shortcut layers in the network exhibits an extreme bimodal characteristic. If traditional linear quantization strategies are used to compress it to INT8, it introduces exponentially amplified quantization errors, resulting in completely distorted power grid data generated during the simulation.

[0005] Finally, traditional model distillation struggles to capture critical "extreme" operating conditions and heavily relies on massive amounts of data. In power grid scenarios, the transmission cost of massive amounts of high-frequency, multimodal data is extremely high; moreover, the traditional distillation loss function uses global average error, which causes the model to smooth out the most critical "extreme risk conditions" in the power grid system (such as overload exceeding limits and sudden short circuits) as noise, resulting in the lightweight model losing its ability to predict and predict power grid disasters.

[0006] Therefore, there is an urgent need to provide a new lightweight processing method for diffusion models that can break through hardware barriers, decouple "high-dimensional extraction by cloud GPUs" from "low-power inference by edge NPUs", and accurately retain the diffusion model's ability to infer high-dimensional conditions of the power grid under extreme conditions, so as to meet the high-efficiency operation requirements of the regional power grid. Summary of the Invention

[0007] To address the technical problems existing in the prior art, this invention provides a method and system for optimizing regional power grid operation based on a multimodal diffusion model. This method fully considers the complex activation distribution within the diffusion model, balances the limitations of quantization compression with the safety red line requirements for power grid disaster prediction, and achieves low-cost transmission of high-dimensional knowledge from multiple dimensions.

[0008] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A regional power grid operation optimization method based on a multimodal diffusion model includes the following steps: S1. Construct a multimodal operation information feature library for the power grid in the target area; the feature library covers power characteristic data reflecting the physical laws of power grid operation and meteorological characteristic data reflecting environmental influencing factors; based on the feature library, a teacher diffusion model with a parameter scale of the first scale is pre-trained on a cloud graphics processor GPU computing cluster, and a student diffusion model with a parameter scale of the second scale is initialized and constructed simultaneously, the second scale being smaller than the first scale; S2. A hybrid sample feature pool covering multiple time steps is constructed using a dynamic update strategy. Based on the hybrid sample feature pool, the student diffusion model is subjected to extreme value-aware dynamic iterative knowledge distillation for the denoising process through the teacher diffusion model, resulting in the distilled student diffusion model. In the backpropagation process, the extreme value-aware dynamic iterative knowledge distillation evaluates the power grid physical state mapped by the current denoised sample in real time. When the sample features indicate that the power grid is in a high-risk extreme value range, the loss weight factor with nonlinear amplification is dynamically calculated and assigned to the sample to amplify the reconstruction error corresponding to the extreme risk conditions. S3. Perform time-step-aware low-bit quantization on the distilled Student diffusion model; construct a calibration dataset based on uniform sampling of the time-step distribution of the Student diffusion model, and execute a separation quantization strategy on the Student diffusion model to generate a quantized diffusion model adapted to the underlying fixed-point operator. S4. Construct a progressive training objective, fine-tune the quantized diffusion model with a very small learning rate in a floating-point computing environment, and convert the lightweight diffusion model obtained after fine-tuning into a static computation graph and distribute it to the edge terminal node equipped with a neural network processor (NPU) in the target area power grid field. S5. Input the real-time collected regional power grid current meteorological data and grid power operation data into the lightweight diffusion model. Perform multi-time step situation simulation on NPU hardware through the lightweight diffusion model to generate power grid situation simulation results within the prediction period. Use the extreme value risk probability and equipment over-limit nodes in the power grid situation simulation results as optimization indicators to adjust the regional power grid operation mode.

[0009] Preferably, in step S2, the extreme value-aware dynamic iterative knowledge distillation sets a weighting factor based on the extreme value distribution of the power grid multimodal data, and guides the parameter update of the student diffusion model by extracting the hidden layer prediction noise and variational lower bound output distribution of the teacher diffusion model at different denoising time steps.

[0010] Preferably, the overall objective function of extreme value-aware dynamic iterative knowledge distillation is:

[0011] in, Losses due to overall objectives To predict noise alignment loss, For variational lower bound alignment loss, This is the balance coefficient.

[0012] Preferably, in step S2, the construction of the mixed sample feature pool specifically includes: sampling initial noise from a standard Gaussian distribution to construct an initial set; and performing a random number of denoising operations on the samples in the initial set using a teacher diffusion model to obtain an extended batch dataset containing mixed samples from different time steps. Expand batch dataset This constitutes a mixed sample feature pool.

[0013] Preferably, in step S3, the time-step-aware low-bit quantization processing specifically includes: establishing a sampling pool covering the total diffusion time steps; uniformly extracting corresponding time steps from the total diffusion time steps according to a preset fixed time step interval; for the extracted corresponding time steps, randomly generating intermediate input data using the Student diffusion model, and summarizing the intermediate input data to form a calibration dataset for activation value quantization calibration. Based on the calibration dataset Calculate the dynamic truncation range of activation values ​​for each network layer to determine the scaling factor and offset for quantization.

[0014] Preferably, in step S3, the execution of the separation quantization strategy specifically includes: identifying the shortcut connection in the student diffusion model used to connect deep features and shallow features; performing independent activation value quantization and weight graph quantization calculations on deep features and shallow features respectively to obtain the corresponding quantization results, and then inputting the corresponding quantization results into the shortcut connection for concatenation operation in the channel dimension.

[0015] Preferably, in step S4, a progressive training objective is constructed based on the standard diffusion loss and the data reconstruction loss, specifically as follows:

[0016] in, Total loss; For standard diffusion loss, For data reconstruction loss, This is a balance parameter that dynamically decays with the number of iterations.

[0017] Preferably, in step S1, the student diffusion model reduces the number of channels in each feature layer by a preset ratio while maintaining the same network depth and topological receptive field as the teacher diffusion model, so as to reduce the theoretical computational complexity of the model. The parameter size of the student diffusion model after reducing the number of channels is one-quarter of that of the teacher diffusion model, and it is isomorphic to the teacher diffusion model using the U-Net architecture.

[0018] Preferably, in step S1, the power characteristic data includes at least one of bus voltage, line power flow and switch status; the meteorological characteristic data includes at least one of ambient temperature, ambient humidity, wind speed and irradiance.

[0019] The present invention also discloses a regional power grid operation optimization system based on a multimodal diffusion model, comprising an interconnected memory and a processor, wherein the memory stores a computer program, and the computer program executes the steps of the method described above when run by the processor.

[0020] Compared with the prior art, the advantages of the present invention are as follows: This invention provides a method that perfectly decouples the physical architecture of high-performance cloud training from low-power edge inference. Through channel pruning and INT8 decoupling quantization, the size of the full-precision multimodal diffusion model can be significantly compressed according to a preset ratio, perfectly matching the underlying operators of the power grid edge NPU, and significantly reducing model inference latency and memory usage. This invention overcomes the limitations of data privacy protection and massive transmission, extracting teacher model knowledge without original data through dynamic iterative distillation; simultaneously, it innovatively introduces an "extreme value weighting mechanism," effectively overcoming the problem of inference distortion in conventional lightweight models when encountering extreme weather or faults, maintaining high fidelity of the generated inference results while ensuring a high compression ratio. This invention provides a forward-looking situational awareness means for decision-making on high-risk sensitive lines and guiding differentiated power grid scheduling, significantly improving the power grid's disaster prevention and mitigation capabilities. Attached Figure Description

[0021] Figure 1 This is a flowchart of a regional power grid operation optimization method based on a multimodal diffusion model, according to an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, the regional power grid operation optimization method based on a multimodal diffusion model provided in this embodiment of the invention relies on a heterogeneous physical architecture of "cloud GPU data center + edge NPU intelligent terminal" and includes the following steps: S1. Construct a multimodal operation information feature library for the target area power grid; the feature library covers two types of modal data: one type is power characteristic data reflecting the physical laws of power grid operation, including at least one of bus voltage, line power flow and switch status; the other type is meteorological characteristic data reflecting environmental influencing factors, including at least one of ambient temperature, ambient humidity, wind speed and irradiance.

[0024] Based on a multimodal runtime information feature library, a teacher diffusion model (Teacher U-Net, FP32, approximately 1.2 Billion parameters) with a first-scale parameter size was pre-trained on a general-purpose graphics processing unit (GPU) computing cluster. Simultaneously, a student diffusion model (Student U-Net, FP32, approximately 0.3 Billion parameters) with a second-scale parameter size was initialized and constructed, where the second scale is smaller than the first. The student diffusion model, while maintaining the same network depth and topological receptive field as the teacher diffusion model, reduces the number of channels in each feature layer by a preset ratio (e.g., Channels / 2) to reduce the theoretical computational complexity (FLOPs). The parameter size of the student diffusion model after reducing the number of channels is approximately one-quarter of that of the teacher diffusion model, and it is isomorphically identical to the teacher diffusion model, adopting the U-Net architecture.

[0025] S2. Abandoning the reliance on massive amounts of target domain data, and leveraging the generative capabilities of the teacher diffusion model, a dynamic update strategy is employed to construct a mixed sample feature pool covering multiple time steps. The construction of this mixed sample feature pool specifically includes: sampling initial noise from a standard Gaussian distribution to construct an initial set; performing a random-step denoising operation (Shuffle Denoise) on the samples in the initial set using the teacher diffusion model to obtain an extended batch dataset containing mixed samples from different time steps. This process yields a mixed sample feature pool. In each training iteration, a sub-batch is randomly selected from the extended batch dataset to train the student diffusion model, and the corresponding samples in the original dataset are replaced with denoised samples to achieve dynamic updates without real data dependencies. This mechanism enables data-free knowledge transfer, effectively mitigating the risks of massive transmission and privacy leaks of sensitive power grid data.

[0026] Based on the aforementioned mixed-sample feature pool, an extreme-value-aware dynamic iterative knowledge distillation method is used to guide the denoising process of the student diffusion model through the teacher diffusion model. Specifically, the extreme-value-aware dynamic iterative knowledge distillation sets weighting factors based on the extreme value distribution of the power grid multimodal data. By extracting the hidden layer prediction noise and variational lower bound output distribution of the teacher diffusion model at different denoising time steps, it guides the parameter updates of the student diffusion model. The overall objective function of the extreme-value-aware dynamic iterative knowledge distillation is... Represented as:

[0027] in, To predict noise alignment loss, For variational lower bound alignment loss, This is the balance coefficient.

[0028] For a given denoising time step With denoised samples Predicting noise alignment loss The calculation formula is:

[0029] in, This indicates an extended batch dataset. All time steps are sample pairs Take the expected value; This is the predicted noise output of the teacher diffusion model. This is the predicted noise output for the student diffusion model; The weighting factor is set based on the multimodal extreme values ​​of the power grid; This is an extended batch dataset for a dynamic iterative distillation process.

[0030] Variational lower bound alignment loss The KL divergence is obtained by calculating the relative entropy (KL divergence) between the teacher diffusion model and the student diffusion model during the back diffusion process. The calculation formula is as follows:

[0031] in, express divergence, and The teacher diffusion model and the student diffusion model are respectively represented by time steps. Stepping into time The inferred probability distribution.

[0032] Among them weighting factors Based on the sample The corresponding multimodal power grid data characteristics are dynamically calculated. During backpropagation, the physical state of the power grid mapped by the current denoised sample is evaluated in real time: when the sample When the corresponding meteorological characteristic prediction value exceeds the preset meteorological disaster threshold, or the power load prediction value exceeds the preset safety margin boundary, the sample is determined to be an extreme risk condition in the power grid situation simulation; at this time, the weighting factor... The overall objective function increases non-linearly. During backpropagation, the distillation error under extreme risk conditions is amplified and penalized, forcing the student diffusion model to prioritize fitting the situational trajectory of extreme power grid disasters while simplifying parameters. When the sample does not fall into the high-risk extreme value range, the weighting factor... Maintain the base weight value.

[0033] This invention innovatively introduces an extreme value weighting mechanism during the distillation process to evaluate the physical state of the power grid mapped by the current denoised sample in real time. When the sample characteristics indicate that the power grid is in a high-risk extreme value range (such as extreme weather conditions, equipment overload, or voltage drop risk), a loss weight factor with nonlinear amplification is dynamically calculated and assigned to the sample. By amplifying the reconstruction error of such high-risk conditions in backpropagation, the student diffusion model is forced to prioritize fitting the situational projection trajectory for extreme power grid disasters while simplifying parameters.

[0034] S3. Perform time-step-aware low-bit quantization on the distilled student diffusion model; construct a calibration dataset based on uniform sampling of the time-step distribution of the diffusion model, and execute a separation quantization strategy in the feature concatenation network layer of the student diffusion model to generate a quantized diffusion model adapted to the underlying fixed-point operators. This step solves the problem of adapting the underlying fixed-point operators when deploying the lightweight model to the NPU, and performs INT8 post-training quantization (PTQ) on the distilled student diffusion model.

[0035] To address the drastic fluctuations in activation distribution caused by multiple time steps in the diffusion denoising process, time-step-aware low-bit quantization is employed. Specifically, this involves establishing a quantization process covering the total diffusion time steps. The sampling pool; according to a preset fixed time step interval. From the total diffusion time step The corresponding time steps are uniformly extracted; for the extracted corresponding time steps, intermediate input data are randomly generated using the Student diffusion model, and the intermediate input data are aggregated to form a calibration dataset for activation value quantization calibration. Based on the calibration dataset The dynamic cutoff range of activation values ​​for each network layer is calculated to determine the scaling factor and offset for INT8 quantization. This calibration dataset D covers the activation distribution throughout the entire denoising lifecycle, avoiding representativeness bias caused by calibration at a single time step.

[0036] To address the quantization error accumulation problem caused by the extreme bimodal activation distribution of the feature concatenation network layer in the student diffusion model, a split shortcut quantization strategy is adopted: the student diffusion model uses a U-Net architecture, and the feature concatenation network layer is used to connect deep features and shallow features; before performing channel concatenation, the deep features are... shallow features Decouple and calculate separately and The dynamic truncation boundary is determined, and independent quantization scaling factors and offsets are obtained. These are then mapped using INT8 before a channel-level concatenation operation is performed. This separate quantization function... Represented as:

[0037] in, This indicates the original, complete features that were not spliced ​​together. This indicates the INT8 discrete quantization operation. This represents the concatenation operation of features along the channel dimension. Deep features shallow features The independent INT8 quantization results are combined into a complete quantization feature through channel concatenation. This mechanism can eliminate the accuracy collapse caused by shallow, extremely large activation values ​​masking deep, fine-grained features.

[0038] S4. Based on standard diffusion loss With data reconstruction loss A progressive training objective is constructed, and the quantized diffusion model is fine-tuned with a very small learning rate in a floating-point computing environment to compensate for the slight feature loss caused by low bit truncation. Finally, the fine-tuned lightweight diffusion model is converted into a static computation graph and distributed, compiled and deployed on the edge terminal nodes of the target area power grid equipped with neural network processors (NPUs) to perform low-latency power grid situation simulation tasks.

[0039] The progressive training objective is:

[0040] in, Total loss; For standard diffusion loss, For data reconstruction loss, This is a balance parameter that dynamically decays with the number of iterations.

[0041] S5. Input the real-time collected regional power grid current meteorological data and grid power operation data into the lightweight diffusion model; perform multi-time-step situation simulation on the NPU hardware through the lightweight diffusion model to generate the power grid situation simulation results within the prediction period; use the extreme value risk probability and equipment over-limit nodes in the power grid situation simulation results as intelligent dispatch indicators to dynamically adjust the start-up and shutdown of power generation equipment or line load balancing strategies of the regional power grid to reduce the comprehensive operation risk value of the regional power grid; after adjustment, input the updated grid power operation data into the lightweight diffusion model again to re-perform situation simulation and risk assessment, forming a closed-loop optimization of the comprehensive operation risk of the regional power grid.

[0042] This invention provides a method that perfectly decouples the physical architecture of high-performance cloud training from low-power edge inference. Through channel pruning and INT8 decoupling quantization, the size of the full-precision multimodal diffusion model can be significantly compressed according to a preset ratio, perfectly matching the underlying operators of the power grid edge NPU, and significantly reducing model inference latency and memory usage. This invention overcomes the limitations of data privacy protection and massive transmission, extracting teacher model knowledge without original data through dynamic iterative distillation; simultaneously, it innovatively introduces an "extreme value weighting mechanism," effectively overcoming the problem of inference distortion in conventional lightweight models when encountering extreme weather or faults, maintaining high fidelity of the generated inference results while ensuring a high compression ratio. This invention provides a forward-looking situational awareness means for decision-making on high-risk sensitive lines and guiding differentiated power grid scheduling, significantly improving the power grid's disaster prevention and mitigation capabilities.

[0043] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0044] Step S1: Construction of heterogeneous diffusion model and multimodal pre-training.

[0045] In a cloud data center, multimodal historical data of the regional power grid are collected, including meteorological data sequences and grid structure characteristic data. A large teacher diffusion model (Teacher U-Net) with full-precision floating-point operations (FP32) is pre-trained using a general-purpose GPU computing cluster. Simultaneously, a simplified student diffusion model (Student U-Net) is initialized. The simplification strategy involves reducing the number of feature channels by a predetermined ratio (Channels / 2) and reducing the number of stacked layers for deep features, while maintaining the same spatial resolution, to construct a lightweight network skeleton.

[0046] Step S2: Dynamic iterative data-free distillation with multi-level extremum awareness.

[0047] Since the original multimodal data is extremely large, this embodiment of the invention uses a dynamic update strategy on the GPU to construct the distillation data: S201. Generate batch by sampling initial noise from a standard Gaussian distribution. ; S202. Perform a denoising operation with random steps on the teacher model to obtain a mixed batch with a uniform time step distribution. ; S203. In each training iteration, from A random sub-batch is selected to train the student model, and the prediction noise alignment loss is calculated. Loss aligned with variational lower bound The samples that have undergone one step of denoising are used to replace the old samples in the pool.

[0048] Extreme value weighting: in calculation At that time, the system performs extreme value determination on the potential semantics contained in the current denoised samples. For example, if the sample features reflect the potential high-risk fault risk or extreme weather conditions of the power grid nodes, the system dynamically calculates and assigns them a higher weighting factor. Gradient backpropagation containing this factor forces the student model to prioritize fitting such high-risk trajectories.

[0049] Step S3: Time step perception and shortcut separation quantization.

[0050] Perform INT8 transformation (PTQ) on the distilled FP32 student model on the hybrid computing platform: S301. Constructing a calibration set: Setting the preset time step interval Multiple corresponding time step nodes are uniformly extracted from the entire diffusion process, and several intermediate features are generated using forward inference of the student model as a calibration dataset. This is to cover the activation distribution throughout the entire lifecycle of denoising.

[0051] S302. Split Shortcut Quantization: This applies to the shortcut layer connecting the Encoder and Decoder in the U-Net model. It separates deep features. shallow features Calculate separately and The dynamic clipping range is calculated by determining the independent quantization scaling factor and offset, performing INT8 mapping on each, and then executing the channel-level quantization. The (Concat) operation eliminates the accumulation of quantization errors caused by the bimodal distribution.

[0052] Step S4: Progressive joint tuning and NPU deployment.

[0053] Introducing a balance parameter that dynamically decays with the number of iterations Construct the overall loss The INT8 model was fine-tuned with a low learning rate. Finally, the model was exported as a static computation graph and distributed to intelligent edge devices equipped with NPUs within the regional power grid substations.

[0054] S5. Regional power grid operation optimization The system collects real-time meteorological forecast data and current network topology within a preset forecast period in the region, and inputs this data into a lightweight diffusion model in the NPU at the substation edge. The model rapidly performs inverse diffusion simulations in parallel, outputting a risk map of future limits for each node and a high-load early warning heat map.

[0055] The dispatching system uses the extreme values ​​of high-risk nodes in the heat map as optimization targets, automatically adjusts the on / off status of regional interconnection switches or executes load transfer strategies, so that the comprehensive risk value of the system re-evaluated by the situation simulation model is reduced to the minimum, forming a closed-loop optimization.

[0056] To verify the effectiveness of the method described in this invention, the following verification process is designed: 1. Experimental Environment and Dataset Construction Historical multimodal operation data of a typical regional power grid were selected as the research object. High-performance GPU nodes were configured in the cloud for teacher model training and extreme value-aware distillation; edge computing gateways equipped with general-purpose NPU chips were configured at the edge for inference and verification of the quantization model.

[0057] 2. Setting of evaluation indicators The overall performance of this invention in model compression and situation projection is verified by comparing the following indicators: Model complexity metrics: Parameter size (Parameters), model storage volume (MB), and peak edge device memory usage.

[0058] Indicators of simulation timeliness: single-step denoising reasoning delay and total time spent on the entire situation simulation.

[0059] The simulation fidelity index is evaluated by measuring the difference in distribution between generated data and real physical data. Emphasis is placed on comparing the accuracy differences between normal stable operating conditions and extreme high-risk operating conditions (such as simulating severe failures or extreme weather).

[0060] 3. Comparison Methods and Verification of Expected Results The method described in this invention was compared with the "original cloud-based FP32 large model" and the "conventional global linear quantization INT8 model" in a comparative experiment. Specific expected comparisons are shown in Tables 1 and 2. Table 1: Comparison and Validation Table of Expected Model Resource Consumption at Each Stage

[0061] Table 2: Comparison and Verification Table of Simulation Speed ​​and Fidelity under Different Operating Conditions

[0062] Based on the theoretical derivation of the aforementioned evaluation mechanism, it is evident that if conventional linear quantization is used for direct compression, under extreme and high-risk operating conditions, the resulting power grid situation is highly susceptible to distortion due to the neglect of extreme value distribution mutations in multimodal data. This invention, however, combines extreme value-aware weighting with a split-shortcut mechanism. Experimentally, this ensures that it achieves extremely low inference latency comparable to conventional INT8 quantization while significantly mitigating accuracy loss under extreme conditions, enabling the lightweight model to achieve inference fidelity highly close to that of large cloud-based models.

[0063] In summary, this invention fully considers the complex activation distribution within the diffusion model, and comprehensively balances the limitations of quantization compression with the safety red line requirements for power grid disaster prediction, thus achieving low-cost transmission of high-dimensional knowledge from multiple dimensions.

[0064] This invention constructs heterogeneous teacher and student diffusion models with varying parameter scales using a cloud-based GPU cluster. It employs a dynamically iterative multi-layer extremum-aware knowledge distillation technique, enabling the student diffusion model to accurately learn the denoised distribution of the teacher model without relying on massive amounts of data. Subsequently, the distilled student diffusion model undergoes extreme compression using an INT8 quantization strategy that separates time-step awareness and shortcut layers. Finally, through progressive joint training and fine-tuning, the lightweight model is deployed to the NPU terminal at the power grid edge. This invention significantly reduces the model size, overcoming the computational bottleneck of deploying large diffusion models on underlying hardware while effectively preserving the model's fidelity in predicting power grid weather and extreme power risk conditions, thus meeting the efficient and intelligent operation requirements of new regional power grids.

[0065] This invention overcomes the technical shortcomings of existing large diffusion models when applied to front-line power grid operations, such as the large number of parameters making them difficult to deploy on edge NPUs, the tendency of conventional quantization compression to cause accuracy collapse, and the tendency of traditional distillation methods to smooth out extreme power grid conditions. By using multi-layer extreme value sensing dynamic iterative distillation and time step sensing shortcut separation quantization, high-fidelity extreme compression of large models to edge terminals is achieved under the premise of physical hardware decoupling.

[0066] This invention also discloses a regional power grid operation optimization system based on a multimodal diffusion model, comprising an interconnected memory and a processor. The memory stores a computer program, which, when run by the processor, executes the steps of the method described above. The system of this invention, corresponding to the method described above, also possesses the advantages described above.

[0067] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0068] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing the operation of a regional power grid based on a multimodal diffusion model, characterized in that, Includes the following steps: S1. Construct a multimodal operation information feature library for the power grid in the target area; the feature library covers power characteristic data reflecting the physical laws of power grid operation and meteorological characteristic data reflecting environmental influencing factors; based on the feature library, a teacher diffusion model with a parameter scale of the first scale is pre-trained on a cloud graphics processor GPU computing cluster, and a student diffusion model with a parameter scale of the second scale is initialized and constructed simultaneously, the second scale being smaller than the first scale; S2. A hybrid sample feature pool covering multiple time steps is constructed using a dynamic update strategy. Based on the hybrid sample feature pool, the student diffusion model is subjected to extreme value-aware dynamic iterative knowledge distillation for the denoising process through the teacher diffusion model, resulting in the distilled student diffusion model. In the backpropagation process, the extreme value-aware dynamic iterative knowledge distillation evaluates the power grid physical state mapped by the current denoised sample in real time. When the sample features indicate that the power grid is in a high-risk extreme value range, the loss weight factor with nonlinear amplification is dynamically calculated and assigned to the sample to amplify the reconstruction error corresponding to the extreme risk conditions. S3. Perform time-step-aware low-bit quantization on the distilled Student diffusion model; A calibration dataset is constructed based on uniform sampling of the time step distribution of the student diffusion model, and a separation quantization strategy is applied to the student diffusion model to generate a quantized diffusion model adapted to the underlying fixed-point operators. S4. Construct a progressive training objective, fine-tune the quantized diffusion model with a very small learning rate in a floating-point computing environment, and convert the lightweight diffusion model obtained after fine-tuning into a static computation graph and distribute it to the edge terminal node equipped with a neural network processor (NPU) in the target area power grid field. S5. Input the real-time collected regional power grid current meteorological data and grid power operation data into the lightweight diffusion model. Perform multi-time step situation simulation on NPU hardware through the lightweight diffusion model to generate power grid situation simulation results within the prediction period. Use the extreme value risk probability and equipment over-limit nodes in the power grid situation simulation results as optimization indicators to adjust the regional power grid operation mode.

2. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 1, characterized in that, In step S2, the extreme value-aware dynamic iterative knowledge distillation sets a weighting factor based on the extreme value distribution of the power grid multimodal data. By extracting the hidden layer prediction noise and variational lower bound output distribution of the teacher diffusion model at different denoising time steps, it guides the parameter update of the student diffusion model.

3. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 2, characterized in that, The overall objective function of extremum-aware dynamic iterative knowledge distillation is: in, Losses due to overall objectives To predict noise alignment loss, For variational lower bound alignment loss, This is the balance coefficient.

4. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 1, 2, or 3, characterized in that, In step S2, the construction of the mixed sample feature pool specifically includes: sampling initial noise from a standard Gaussian distribution to construct an initial set; and performing a random number of denoising operations on the samples in the initial set using a teacher diffusion model to obtain an extended batch dataset containing mixed samples from different time steps. Expand batch dataset This constitutes a mixed sample feature pool.

5. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 1, 2, or 3, characterized in that, In step S3, the time-step-aware low-bit quantization processing specifically includes: establishing a sampling pool covering the total diffusion time steps; uniformly extracting corresponding time steps from the total diffusion time steps according to a preset fixed time step interval; for the extracted corresponding time steps, randomly generating intermediate input data using the Student diffusion model, and summarizing the intermediate input data to form a calibration dataset for activation value quantization calibration. Based on the calibration dataset Calculate the dynamic truncation range of activation values ​​for each network layer to determine the scaling factor and offset for quantization.

6. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 1, 2, or 3, characterized in that, In step S3, the execution of the separation quantization strategy specifically includes: identifying the shortcut connection used to connect deep features and shallow features in the student diffusion model; performing independent activation value quantization and weight graph quantization calculations on deep features and shallow features respectively to obtain the corresponding quantization results; and then inputting the corresponding quantization results into the shortcut connection for concatenation operation in the channel dimension.

7. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 1, 2, or 3, characterized in that, In step S4, a progressive training objective is constructed based on the standard diffusion loss and the data reconstruction loss, specifically as follows: in, Total loss; For standard diffusion loss, For data reconstruction loss, This is a balance parameter that dynamically decays with the number of iterations.

8. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 1, 2, or 3, characterized in that, In step S1, the student diffusion model reduces the number of channels in each feature layer by a preset ratio while maintaining the same network depth and topological receptive field as the teacher diffusion model, so as to reduce the theoretical computational complexity of the model. The parameter size of the student diffusion model after reducing the number of channels is one-quarter of that of the teacher diffusion model, and it is isomorphic to the teacher diffusion model using the U-Net architecture.

9. The regional power grid operation optimization method based on a multimodal diffusion model according to claim 1, 2, or 3, characterized in that, In step S1, the power characteristic data includes at least one of bus voltage, line power flow and switch status; the meteorological characteristic data includes at least one of ambient temperature, ambient humidity, wind speed and irradiance.

10. A regional power grid operation optimization system based on a multimodal diffusion model, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-9.

Images