Power distribution network situation deduction method and system considering multi-type source-load interaction

Abstract

The invention discloses a power distribution network situation deduction method and system considering multi-type source-load interaction, and relates to the technical field of power distribution network situation prediction. The method comprises the following steps: acquiring operation states of various types of source loads and power distribution equipment in a power distribution network; capturing a collaborative prediction relationship between the source load and the power distribution equipment by using a multivariable time sequence coupling model; constructing an electric power AI large model, and pre-training the electric power AI large model by using an autoregression method; performing parameter fine tuning on the pre-trained electric power AI large model by adopting a plurality of fine tuning technologies; and deducing the situation of the power distribution network according to the operation states of the source load and the power distribution equipment and the collaborative prediction relationship between the source load and the power distribution equipment by using the fine-tuned power AI large model. According to the method, on the premise that large-scale flexible resources are contained and the flexible resources continuously respond to the interaction strategy of the power distribution network, calculation intelligent deduction of multi-type source-load interaction of the short-term operation situation of the power distribution network can be achieved.

Description

Technical Field

[0001] This invention relates to the field of distribution network situation prediction technology, and in particular to a distribution network situation prediction method and system that considers the interaction of multiple types of sources and loads. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] As the scale of flexible resources connected to the distribution network continues to increase and the depth of interaction between the distribution network's sources, grid, load, and storage deepens, the operational characteristics of the new distribution network are significantly different from those of the traditional distribution network. Modern new distribution networks have become ultra-high-dimensional, complex, nonlinear systems where information and physical networks are intertwined and their coupling is constantly deepening. With the increasing coupling of information flow and power flow, the resulting changes in system characteristics are becoming more and more pronounced and complex.

[0004] Currently, there is limited research on distributed intelligent computing and simulation technology for multi-temporal and spatial situational awareness in distribution networks that considers various types of source-load interactions. Research primarily focuses on the processing of massive and complex system data, but it still has the following shortcomings: (1) Insufficient system modeling dimensions. Most existing studies still regard the distribution network as a physical network dominated by "power flow alone", and lack a unified descriptive framework for the ultra-high dimensionality, nonlinearity and strong time-varying characteristics generated by the coupling of information and physical systems. The response behavior of flexible resources (distributed wind power, distributed / centralized photovoltaic, energy storage, flexible loads) and the feedback loop between communication delay, measurement error and control commands have not been fully characterized, resulting in a sharp drop in model accuracy when the spatiotemporal scale is expanded.

[0005] (2) Poor adaptability of distributed computing architecture. Existing mainstream solutions use general MapReduce or GPU-graph computing frameworks, treating the spatiotemporal coupling power flow of the distribution network as a regular large graph or batch matrix operation, ignoring the need for millisecond-second real-time synchronization for power balancing, and the requirement for minute-hour rolling deduction for state estimation and optimization, which is a "multi-clock frequency" requirement. As a result, two extremes occur: either accuracy is sacrificed for real-time performance, or real-time performance is sacrificed for scale, making it difficult to achieve both. In addition, the contradiction between the storage capacity and communication bandwidth of embedded platforms and the exponential growth of the number of distribution network nodes is becoming increasingly acute. Existing partitioning strategies and sliding window algorithms generate a large amount of redundant synchronization and load imbalance in scenarios of dynamic node deployment and withdrawal and frequent topology switching.

[0006] (3) Incomplete situation prediction closed loop. The existing "multi-timescale state estimation + model predictive control" link only solves the one-way process from state perception to optimized scheduling, lacking closed-loop verification from scheduling instructions to flexible resource response and then to network state re-perception. When distributed photovoltaic, energy storage and flexible loads are continuously adjusted according to the interactive strategy, their random, nonlinear and dispersed responses will introduce new uncertainties, causing the original state estimation error to be amplified and the predicted trajectory to deviate rapidly from the actual operating point. At present, there is no mechanism that can correct this deviation online, in a distributed and adaptive manner, making it difficult to guarantee the consistency of decision-making between medium- and long-term planning and short-term scheduling.

[0007] In summary, existing distributed intelligent computing and simulation models for distribution networks with multi-temporal and spatial scales that consider the interaction of various types of sources and loads have problems such as insufficient coupling capability of multi-type source and load data, inability to balance computational timeliness and accuracy, and lack of correction mechanisms. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide a distribution network situation prediction method and system that considers the interaction of multiple types of sources and loads. Under the premise of having large-scale flexible resources and the flexible resources continuously responding to the distribution network interaction strategy, it can realize the intelligent prediction of the short-term operation status of the distribution network by multi-type source and load interaction, give full play to the role of multi-energy synergy, and provide important guidance for distribution network operation scheduling and medium- and long-term planning.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a method for power distribution network situation prediction considering multi-type source-load interactions, comprising the following steps: To obtain the operating status of various types of source loads and power distribution equipment in the power distribution network; A multivariate time series coupling model is used to capture the collaborative prediction relationship between source load and power distribution equipment; Construct a large-scale AI model for the power industry and pre-train it using an autoregressive method; A combination of fine-tuning techniques was used to fine-tune the parameters of the pre-trained power AI model. The finely tuned power AI model is used to extrapolate the power distribution network situation based on the operating status of power sources and distribution equipment, as well as the collaborative prediction relationship between power sources and distribution equipment.

[0010] Furthermore, the operating status of power sources and distribution equipment includes the operating status vectors of distributed power sources, smart loads, energy storage devices, and distribution network equipment.

[0011] Furthermore, when using a multivariate time series coupling model to capture the collaborative prediction relationship between source loads and distribution equipment, a dynamic coupling matrix is ​​used to quantify the interaction relationship between source loads, and a system of differential equations is used to describe the evolution process of the distribution network situation.

[0012] Furthermore, the power AI big data model is a model with billions of parameters.

[0013] Furthermore, the specific steps for constructing a large-scale AI model for the power industry and pre-training it using an autoregressive method are as follows: A large-scale AI model for the power industry was built based on deep learning technology and the Transformer architecture. The objective function of the autoregressive method is constructed with the goal of maximizing the conditional probability of future states given historical states. Based on the objective function, an autoencoder algorithm is used to pre-train a large-scale AI model for the power industry.

[0014] Furthermore, the large-scale AI model for power systems includes an input embedding layer, an encoder module, a decoder module, a multi-head attention mechanism module, a feedforward neural network module, and an output projection layer.

[0015] Furthermore, various fine-tuning techniques, including Full Fine-tuning, LoRA, and Prompt-tuning, are used to perform full parameter updates, low-rank decomposition, and soft hint optimization on the large power AI model, respectively.

[0016] A second aspect of the present invention provides a distribution network situation simulation system considering multi-type source-load interactions, comprising: The data acquisition module is configured to acquire the operating status of various types of source loads and power distribution equipment in the power distribution network; The collaborative relationship prediction module is configured to capture the collaborative prediction relationship between source load and power distribution equipment using a multivariate time series coupling model; The model training module is configured to build a large-scale AI model for the power industry and to pre-train the large-scale AI model for the power industry using an autoregressive method. The fine-tuning module is configured to use a combination of fine-tuning techniques to fine-tune the parameters of the pre-trained power AI large model. The situation simulation module is configured to use a fine-tuned power AI model to simulate the situation of the power distribution network based on the operating status of the power sources and distribution equipment, and the collaborative prediction relationship between the power sources and distribution equipment.

[0017] A third aspect of the present invention provides a computer-readable storage medium storing a computer program adapted for loading by a processor and executing steps in the distribution network situation simulation method considering multiple types of source-load interactions as described in the first aspect of the present invention.

[0018] A fourth aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the distribution network situation simulation method considering multi-type source-load interaction as described in the first aspect of the present invention.

[0019] The above one or more technical solutions have the following beneficial effects: This invention discloses a method and system for power distribution network situation prediction considering the interaction of multiple types of source loads. First, it studies the operating states of various types of source loads and distribution equipment, as well as the collaborative prediction between source loads and distribution equipment, overcoming the problem of insufficient modeling dimensions in existing technologies. Second, based on the above research, a power AI model with billions of parameters is built. Then, autoregressive (AR) pre-training and autoencoder (AE) pre-training are used to train the large model, ensuring computational speed without reducing the model's prediction accuracy. Finally, fine-tuning, parameter-efficient fine-tuning, and prompt-tuning techniques are used to jointly optimize the power AI model, achieving a complete closed loop in the situation prediction process.

[0020] This invention establishes a multi-dimensional state vector modeling system covering distributed power sources, smart loads, energy storage devices, and distribution network equipment, quantifying the mutual influence relationships between devices through a dynamic coupling matrix. It proposes a multivariate time-series coupling model and a situation evolution dynamic equation, using the coupling matrix to quantify the mutual influence relationships between devices and achieve mathematical modeling of complex interactions between source and load devices. Subsequently, a billion-parameter Transformer model architecture specifically designed for distribution network situation simulation is constructed, including a complete design of input embedding, multi-head attention, and feedforward networks. Finally, three fine-tuning techniques—Full Fine-tuning, Parameter-efficient Fine-tuning, and Prompt-tuning—are integrated to fine-tune the pre-trained power AI model, achieving efficient model adaptation from general to specific applications.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the distribution network situation simulation method considering multiple types of source-load interaction in Embodiment 1 of the present invention; Figure 2 This is a structural diagram of the multi-head attention mechanism in Embodiment 1 of the present invention; Figure 3 This is a flowchart of the autoregressive pre-training algorithm in Embodiment 1 of the present invention; Figure 4 This is a flowchart of the autoencoder algorithm in Embodiment 1 of the present invention; Figure 5 This is a structural diagram of the automatic encoder in Embodiment 1 of the present invention; Figure 6 This is a flowchart of the fine-tuning algorithm in Embodiment 1 of the present invention. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] Example 1: Embodiment 1 of the present invention provides a method for power distribution network situation simulation considering multi-type source-load interactions, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the operating status of various types of source loads and power distribution equipment in the power distribution network.

[0027] In one specific implementation, the operating states of power sources and distribution equipment include the operating state vectors of distributed power sources, smart loads, energy storage devices, and distribution network equipment.

[0028] First, the operating states of various types of power sources and equipment in the power distribution network are analyzed and mathematically modeled in depth, and the operating state vector of distributed power sources is defined as follows: (1).

[0029] in, This represents the operating state vector of the distributed power source. This represents the active power output (kW) of the i-th distributed power source; This represents the reactive power output (kVar) of the i-th distributed power source; This represents the terminal voltage amplitude (kV) of the i-th distributed power source; This represents the frequency (Hz) of the i-th distributed power source. This represents the phase angle (°) of the i-th distributed power source; This represents the operating efficiency of the i-th distributed power source.

[0030] Define the operating state vector of the intelligent load as follows: (2).

[0031] in, This is the operating state vector of the intelligent load. This represents the active power demand (kW) of the j-th load. This represents the reactive power demand (kVar) of the j-th load. This represents the response characteristic parameter of the j-th load; This represents the temperature or environmental parameter (°C) of the j-th load. This represents the proportion of the dispatchable capacity of the j-th load.

[0032] Define the operating state vector of the energy storage device as follows: (3).

[0033] in, This is the operating state vector of the energy storage device. This represents the charging and discharging power (kW) of the k-th energy storage device; This represents the state of charge (%) of the k-th energy storage device; This represents the health status (%) of the k-th energy storage device; This represents the temperature (°C) of the k-th energy storage device; This represents the current capacity (kWh) of the k-th energy storage device; This represents the depth of discharge (%) of the k-th energy storage device.

[0034] Define the state vector of the power distribution network equipment as follows: (4).

[0035] in, This represents the state vector of the power distribution network equipment. This represents the current amplitude (A) of the l-th line; This represents the voltage magnitude (kV) at the l-th node; This represents the voltage phase angle (°) at the l-th node; This indicates the state (0 / 1) of the l-th switching device; This indicates the temperature (°C) of the l-th transformer.

[0036] Step 2: Use a multivariate time series coupling model to capture the collaborative prediction relationship between source load and power distribution equipment.

[0037] In one specific implementation, a collaborative prediction relationship between power sources and distribution equipment is established, and the mutual influence and dynamic evolution of the equipment are described using a multivariate time series coupling model. When using the multivariate time series coupling model to capture the collaborative prediction relationship between power sources and distribution equipment, a dynamic coupling matrix is ​​used to quantify the interaction between power sources and power sources, and a system of differential equations is used to describe the evolution process of the power distribution network situation.

[0038] The specific coupling model is as follows: (5).

[0039] in, This represents the system situation prediction vector at time t+Δt; Represents a nonlinear collaborative prediction function; Let represent the operating state vectors of the distributed power source, the smart load, the energy storage device, and the power distribution network device at time t, respectively. The control input vector at time t can be derived from real-time control commands of the distribution network dispatching system, adjustment signals of automated control equipment, dispatching commands of the demand response program, charging and discharging control strategies of the energy storage system, etc. This represents the environmental factor vector at time t, such as temperature, humidity, wind speed, light intensity, electricity price signal, market demand, factors affecting load (holidays, special events, etc.), and operating parameters of the upstream power grid. Δt represents the system's random noise term; Δt represents the prediction time step.

[0040] The interaction between source and load is quantified using a dynamic coupling matrix: (6).

[0041] in, Let represent the coupling matrix at time t; Let represent the influence coefficient of the state of the q-th device at time t on the state of the p-th device; a and b represent the number of affected devices and the number of devices affected, respectively.

[0042] The evolution of the power distribution network situation can be described by the following set of differential equations: (7).

[0043] in, Represents the system state vector; Represents the system dynamic matrix; Represents the control input matrix; Represents the perturbation input matrix; Represents the control input vector; Represents the environmental disturbance vector; This represents the system noise vector.

[0044] In the above formulas, Formula 5 serves as the master prediction model for forecasting the system state at future times. Formula 6 quantifies the mutual influence between devices, providing coupling weight parameters for the F function in Formula 5 and describing the sensitivity between the state variables of different devices. Formula 7 describes the continuous-time dynamic characteristics of the system, providing the theoretical basis for the discretization of Formula 5. It transforms the continuous dynamic equation into the discrete prediction model of Formula 5 through numerical integration.

[0045] In summary, this embodiment uses Equation 5 as the training target, and the AI ​​model learns the mapping relationship of the F function; the coupling relationship of Equation 6 serves as the weight reference for the attention mechanism; and the dynamic characteristics of Equation 7 guide the temporal modeling of the model.

[0046] Step 3: Construct a large-scale AI model for the power industry and pre-train it using an autoregressive method.

[0047] Step 3.1: Build a large-scale AI model for the power industry based on deep learning technology and the Transformer architecture.

[0048] In one specific implementation, different parameter scales of the model are suitable for different scenarios. For example, a parameter scale of millions is suitable for simple distribution network nodes (<100), while a parameter scale of tens of millions is suitable for medium-sized distribution networks (100-1000 nodes). In this embodiment, the power AI large-scale model is a model with billions of parameters. However, a billion-parameter model requires a massive dataset spanning more years, resulting in progressively increasing computational load and training difficulty. Therefore, this embodiment designs a pre-training process based on autoregressive methods and a fine-tuning process based on a combination of multiple technologies to achieve accurate modeling of the billion-parameter model.

[0049] The large-scale AI model for the power industry comprises an input embedding layer, an encoder module, a decoder module, a multi-head attention mechanism module, a feedforward neural network module, and an output projection layer. This embodiment adopts a Transformer architecture for the large-scale AI model, with the multi-head attention mechanism module and the feedforward neural network module embedded as core components in each layer of the encoder and decoder: The input embedding layer is followed by the encoder module, where each layer of the encoder module includes a multi-head attention mechanism module, a feedforward neural network module, a residual connection module, and a layer normalization module. This is then connected to the decoder module, where each layer of the decoder module includes a multi-head attention mechanism module, a feedforward neural network module, a residual connection module, and a layer normalization module. Finally, the output projection layer is connected.

[0050] The overall architecture of the power AI big data model can be defined as follows: (8).

[0051] in, This represents a large-scale AI model function for electricity. Indicates the input embedding layer; Indicates the encoder module, which contains = 48 floors; The decoder module contains = 48 floors; This represents the multi-head attention mechanism module; This represents a feedforward neural network module; Indicates the output projection layer; Represents the set of model parameters. .

[0052] The input embedding layer can convert the states of multiple types of source and load devices into high-dimensional vectors. The multi-head attention mechanism can capture the correlation between different device states. The feedforward neural network module is used for nonlinear feature transformation.

[0053] Specifically, the input embedding layer is represented as: (9).

[0054] in, This indicates the input to the embedded layer. Word embeddings indicating device status; Indicates time location encoding; Indicates the device type code; L is the sequence length. For model dimensions.

[0055] like Figure 2 As shown, the multi-head attention mechanism and each attention head can be represented as follows: (10).

[0056] in, This indicates a multi-head attention mechanism. The attention function is defined as follows: (11).

[0057] in, These represent the query, key, and value matrices, respectively. To learn the parameter matrix; This is the output projection matrix; h = 32 represents the number of attention heads. .

[0058] A feedforward neural network can be represented as: (12).

[0059] in, The first layer weight matrix maps the model dimensions to the hidden dimensions of the feedforward network; The second-layer weight matrix maps the hidden dimensions back to the model dimensions. is the dimension of the hidden layers in the feedforward network; ; This is the bias vector.

[0060] In this embodiment, the scaled dot product attention mechanism is the core computational method of the multi-head attention mechanism, which improves model performance by using multiple scaled dot product attention heads in parallel. Specifically, the basic component of the multi-head attention mechanism is the scaled dot product attention mechanism, such as... Figure 2 As shown in the attached diagram on the left, it first calculates the relevance between the input query, key, and value, then performs scaling and softmax operations to obtain weights, and finally multiplies the weights by the values ​​to obtain the attention output.

[0061] Multi-head attention mechanisms, building upon scaled dot product attention, enhance the model's expressive power by computing multiple attention heads in parallel, such as... Figure 2As shown in the attached diagram on the right, each attention head uses a different linear transformation to compute attention weights through queries, keys, and values, thus capturing different features and relationships in the input.

[0062] Step 3.2: Construct the objective function of the autoregressive method with the goal of maximizing the conditional probability of the future state under a given historical state.

[0063] In one specific implementation, this embodiment employs an autoregressive pre-training method to train a large-scale power AI model, enabling the model to learn the temporal patterns and causal relationships of distribution network situation evolution, such as... Figure 3 As shown, the input data is denoised by the denoising encoder and then the loss function (LOS) is calculated, i.e., the basic autoregressive loss (Equation 13). The weight parameters are updated by backpropagation based on the calculation results, and the LOSS is recalculated using the updated weight parameters, i.e., the multi-step prediction loss function (Equation 16). If convergence is achieved, training ends; otherwise, parameter fine-tuning is performed. The input data of the denoising encoder is a single-time snapshot, which can emphasize the spatial structure and mainly learns the "compression-reconstruction" representation.

[0064] Calculate the basic autoregressive loss. The goal of autoregressive pre-training is to maximize the conditional probability of future states given historical states: (13).

[0065] in, It is the loss function of autoregressive pre-training (the goal is to minimize it). This represents the distribution network state vector at time t; This represents all historical states before time t; θ represents the conditional probability of a future state given a historical state; T represents the length of the training sequence; and θ represents the set of model parameters.

[0066] To ensure the causality of autoregressive training, a causal masking mechanism is employed: (14).

[0067] The attention calculation after masking is as follows: (15).

[0068] In this embodiment, Equation 15 indicates that during the training process of intelligent power distribution network situation simulation, this mechanism strictly controls the direction of information flow by constructing a special mask matrix to prevent the model from acquiring "future information" when predicting future states, thereby violating the temporal causality law. Specifically, when the model processes time series data, the mask matrix sets all attention weights representing information at future moments to negative infinity. After softmax normalization, these weights become zero, achieving complete masking of future information. This design ensures that when the model learns the state evolution laws of power distribution network source and load equipment, it strictly adheres to the physical constraint of "predicting the future only based on historical information," ensuring that the learned prediction patterns are realistically feasible. In the multi-head attention mechanism, each attention head applies the same causal mask, ensuring the consistency of temporal logic during parallel computation. Through this mechanism, the model can learn the real temporal dependencies of state changes of various types of equipment such as distributed power sources, smart loads, and energy storage devices in the power distribution network, providing a reliable causal reasoning basis for subsequent situation prediction.

[0069] To improve long-term forecasting capability, a multi-step forecasting loss function is designed as follows: (16).

[0070] Where H represents the maximum prediction step size; Let the weight coefficients for the h-th step prediction satisfy... ; This represents all states from the beginning to time t.

[0071] The multi-step prediction loss function in Formula 16 effectively addresses the problem of error accumulation and amplification in long-term prediction by simultaneously optimizing the prediction accuracy across multiple future time steps. This function integrates the prediction losses of steps 1, 2, 3, up to H steps through a weighted summation. The weighting coefficients are designed according to the principle of "short-term importance outweighs long-term importance," reflecting the practical need for more accurate predictions in short-term scheduling than long-term planning in distribution network operation. During training, the model needs to learn prediction tasks at multiple time scales simultaneously, forcing the network to develop more robust feature representations and a more stable prediction mechanism. For predicting the state of source and load equipment in distribution networks, the multi-step loss function is particularly suitable for handling dynamic processes with different time scale characteristics, such as distributed power fluctuations, energy storage charging and discharging cycles, and load demand changes.

[0072] Step 3.3: Based on the objective function, pre-train the large-scale power AI model using an autoencoder algorithm. This embodiment employs a dual pre-training strategy: first, an autoregressive method is used to learn temporal causal relationships, and then an autoencoder method is used to learn the intrinsic representation of the state. The two methods are jointly optimized through a joint loss function.

[0073] In one specific implementation, an autoencoder is an unsupervised method for data dimensionality compression and data feature representation, such as... Figure 4 As shown, after inputting the distribution network status data, data preprocessing and normalization operations are performed. The status data is a snapshot of the overall operating status of the distribution network at a specific moment, containing comprehensive operating information of all source and load equipment. It can be obtained through SCADA systems, PMUs, smart meters, operating records stored in data centers, load forecasts, renewable energy power forecasts, meteorological data, electricity price signals, etc.

[0074] Initialize the encoder and decoder parameters, setting hyperparameters such as learning rate and batch size, and then begin training iterations. The training process includes state compression using the encoder to generate latent representation vectors, state reconstruction using the decoder and calculation of reconstruction loss, contrastive learning for reinforcement, calculation of the total loss function, and backpropagation to update parameters based on the total loss function calculation results until the loss converges or the maximum number of rounds is reached. Finally, the performance of the trained model is evaluated using a validation set, and the parameters of the pre-trained model are saved. Input distribution network status data undergoes data preprocessing and normalization. This status data includes multi-dimensional state vectors from distributed power sources, smart loads, energy storage devices, and distribution network equipment. Standardization eliminates dimensional differences between different types of equipment. Initialize the encoder and decoder parameters, setting hyperparameters such as learning rate and batch size, and then begin training iterations.

[0075] During training, the encoder first performs hierarchical compression processing on the states of various source-load devices, encoding the state vectors of distributed power sources, smart loads, energy storage devices, and distribution network devices into corresponding feature representations. Then, it fuses the interaction information between different types of devices through a cross-attention mechanism to generate a latent representation vector containing source-load interaction features. The decoder uses this latent representation vector to reconstruct the states, reconstructing the state parameters of each type of device to ensure that the reconstruction results maintain the coordination relationship between source-load devices.

[0076] The reconstruction loss calculation process employs a classification-weighted strategy, designing weight coefficients based on the varying importance of different types of equipment. The reconstruction loss function is represented as a weighted sum of reconstruction errors for each type of equipment, with distributed power sources and energy storage devices typically assigned higher weights to reflect their crucial role in power balance. The contrastive learning enhancement process improves the model's ability to identify source-load interaction patterns by constructing positive and negative sample pairs. Positive sample pairs select state combinations of different types of equipment at the same time to reflect collaborative operation characteristics, while negative sample pairs select unrelated equipment state combinations. The contrastive learning loss maximizes the similarity of positive samples and minimizes the similarity of negative samples.

[0077] The overall loss function comprehensively considers reconstruction loss, contrastive learning loss, and source-load interaction regularization term, forming a joint optimization objective through weighted combination. The source-load interaction regularization term constrains the latent representation to reflect the true coupling relationships between multiple device types. Parameters are updated via backpropagation based on the overall loss function calculation results. During training, the weights of each loss term are dynamically adjusted to balance different optimization objectives until the loss converges or the maximum number of epochs is reached. Finally, the performance of the trained model is evaluated using a validation set to verify the accuracy of the modeling of multi-type source-load interaction relationships, and the parameters of the pre-trained model are saved.

[0078] Specifically, in most contexts involving autoencoders, the compression and decompression functions are implemented using neural networks. Autoencoder pre-training is employed to learn the latent representation of the distribution network state, improving the model's ability to identify and reconstruct abnormal conditions and rare events. Autoencoder pre-training learns the intrinsic representation of the data by reconstructing the input.

[0079] like Figure 5 As shown, AE can be viewed as a three-layer neural network structure: an input layer, a hidden layer, and an output layer. In terms of data scale, the input layer and the output layer have the same size. Here, N represents the size of the input layer (which also serves as the output layer); M represents the size of the hidden layer. These represent vectors in the input, hidden, and output layers, respectively, and also represent the number of data points in each layer. Here, the number of data points in the hidden layer *h* is less than the number of data points in the input and output layers. and The mapping matrix from the input layer x to the hidden layer h is used to calculate h, and then the mapping matrix from the hidden layer h to the output layer is used to calculate h. . These represent the bias vectors on the hidden layer and the output layer, respectively. The weight matrix between the input layer and the hidden layer, i.e., the mapping matrix from x to h, is an n-m matrix; This represents the weight matrix between the hidden layer and the output layer, i.e., from h to... The mapping matrix is ​​an m x n matrix, which is also the inverse matrix of W.

[0080] Step 4: Use a combination of fine-tuning techniques to fine-tune the parameters of the pre-trained power AI model.

[0081] In one specific implementation, multiple fine-tuning techniques, including Full Fine-tuning, LoRA, and Prompt-tuning, are used for full parameter updates, low-rank decomposition, and soft hint optimization of the large-scale power AI model, respectively. This embodiment employs three different fine-tuning techniques to perform task-specific parameter tuning of the large-scale power AI model, achieving the conversion from a general pre-trained model to a dedicated inference model, such as... Figure 6 As shown, after loading the pre-trained power AI model, the fine-tuning parameters are initialized, and the fine-tuning method is selected. Traditional full-parameter fine-tuning updates all model parameters to adapt to a specific task. Parameter-efficient fine-tuning uses the LoRA (Low-Rank Adaptation) method for efficient parameter fine-tuning, updating only a small number of parameters. Prompt-tuning uses a soft-cue tuning method, optimizing learnable cue vectors to adapt to the distribution network situation prediction task. Afterwards, joint loss calculation is performed; if the performance meets the target, the process ends; otherwise, hyperparameter adjustment is performed.

[0082] The joint optimization of three fine-tuning techniques is a key technical step in realizing the intelligent situational prediction model of distribution networks from general pre-training to dedicated deployment. This scheme adopts a phased and adaptive strategy to coordinate the complementary advantages of Full Fine-tuning, Parameter-efficient Fine-tuning (LoRA), and Prompt-tuning. In practical applications, the selection of fine-tuning methods is mainly based on a comprehensive evaluation of four key factors: computational resource constraints, training data scale, task complexity, and deployment timeliness requirements. When facing large-scale distribution networks (more than 1,000 nodes) with sufficient computing resources, the system prioritizes a comprehensive joint strategy of the three technologies to achieve the best prediction accuracy. For medium-sized distribution networks (100-1,000 nodes), a combination of LoRA and Prompt-tuning is usually chosen to control computational overhead while ensuring performance. For small-scale distribution networks or resource-constrained edge deployment scenarios, Prompt-tuning technology is mainly relied upon for rapid adaptation. This hierarchical selection mechanism is not a simple choice between the three, but a flexible combination based on actual needs, with each technology having its specific applicable scenarios and optimization objectives. In the parallel optimization process using multiple techniques, the system employs a dynamic weight adjustment priority strategy. Initially, Prompt-tuning receives the highest priority because the soft cue mechanism can quickly adapt to new tasks, providing a good starting point for subsequent optimizations. In the mid-stage, the weight of LoRA is gradually increased, finely adjusting the model's representational capabilities through efficient parameter updates via low-rank decomposition. In the final stage, Full Fine-tuning has the highest weight, comprehensively optimizing key parameters to achieve optimal task-specific performance. This phased dynamic weight adjustment mechanism ensures the stability and convergence of the training process, avoiding optimization conflicts that may arise from parallel multi-technique approaches. A strategy combining Bayesian optimization and grid search can be used to adjust hyperparameters. The learning rate uses cosine annealing scheduling, gradually decreasing from an initial 1e-4 to 1e-6; the rank parameter of LoRA is adaptively selected based on model complexity, typically between 16 and 64; the cue length is dynamically adjusted based on task complexity, gradually optimizing from an initial 100 tokens to the optimal length; and the batch size is determined comprehensively based on memory constraints and convergence speed. The entire adjustment process is controlled by an automated hyperparameter optimization algorithm, combined with an early stopping mechanism and model checkpoint saving, to ensure the efficiency and robustness of the training process.

[0083] Step 5: Utilize the finely tuned power AI model to extrapolate the power distribution network situation based on the operating status of power sources and distribution equipment, and the collaborative prediction relationship between power sources and distribution equipment.

[0084] Example 2: Embodiment 2 of the present invention provides a distribution network situation simulation system considering multi-type source-load interactions, including: The data acquisition module is configured to acquire the operating status of various types of source loads and power distribution equipment in the power distribution network.

[0085] In the data acquisition module, the operating status of power sources and distribution equipment includes the operating status vectors of distributed power sources, smart loads, energy storage devices, and distribution network equipment.

[0086] The collaborative relationship prediction module is configured to capture the collaborative prediction relationship between source load and power distribution equipment using a multivariate time series coupling model.

[0087] In the collaborative relationship prediction module, when using a multivariate time series coupling model to capture the collaborative prediction relationship between source load and power distribution equipment, a dynamic coupling matrix is ​​used to quantify the interaction relationship between source load and power distribution equipment, and a set of differential equations is used to describe the evolution process of the power distribution network situation.

[0088] The model training module is configured to build a large-scale AI model for the power industry and pre-train it using an autoregressive method. This large-scale AI model has billions of parameters.

[0089] The model training module is also configured as follows: A large-scale AI model for the power industry is built based on deep learning technology and the Transformer architecture. The large-scale AI model for the power industry includes an input embedding layer, an encoder module, a decoder module, a multi-head attention mechanism module, a feedforward neural network module, and an output projection layer.

[0090] The objective function of the autoregressive method is constructed with the goal of maximizing the conditional probability of future states given historical states.

[0091] Based on the objective function, an autoencoder algorithm is used to pre-train a large-scale AI model for the power industry.

[0092] The fine-tuning module is configured to use a combination of fine-tuning techniques to fine-tune the parameters of the pre-trained power AI model.

[0093] The fine-tuning module employs various fine-tuning techniques, including Full Fine-tuning, LoRA, and Prompt-tuning, which are used to perform full parameter updates, low-rank decomposition, and soft hint optimization on the large power AI model, respectively.

[0094] The situation simulation module is configured to use a fine-tuned power AI model to simulate the situation of the power distribution network based on the operating status of the power sources and distribution equipment, and the collaborative prediction relationship between the power sources and distribution equipment.

[0095] Example 3: Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program adapted for loading by a processor and executing the steps in the distribution network situation simulation method considering multi-type source-load interactions as described in Embodiment 1 of the present invention.

[0096] Example 4: Embodiment 4 of the present invention provides a computer device, the device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps in the distribution network situation simulation method considering multi-type source-load interaction as described in Embodiment 1 of the present invention.

[0097] The steps and methods involved in Examples 2, 3 and 4 above correspond to those in Example 1. For specific implementation details, please refer to the relevant description section of Example 1.

[0098] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc. The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for power distribution network situation simulation considering multi-type source-load interactions, characterized in that, Includes the following steps: To obtain the operating status of various types of source loads and power distribution equipment in the power distribution network; A multivariate time series coupling model is used to capture the collaborative prediction relationship between source load and power distribution equipment; Construct a large-scale AI model for the power industry and pre-train it using an autoregressive method; A combination of fine-tuning techniques was used to fine-tune the parameters of the pre-trained power AI model. The finely tuned power AI model is used to extrapolate the power distribution network situation based on the operating status of power sources and distribution equipment, as well as the collaborative prediction relationship between power sources and distribution equipment.

2. The distribution network situation simulation method considering multi-type source-load interaction as described in claim 1, characterized in that, The operating status of power sources and distribution equipment includes the operating status vectors of distributed power sources, smart loads, energy storage devices, and distribution network equipment.

3. The distribution network situation simulation method considering multi-type source-load interaction as described in claim 1, characterized in that, When using a multivariate time series coupling model to capture the collaborative prediction relationship between source loads and power distribution equipment, a dynamic coupling matrix is ​​used to quantify the interaction between source loads, and a system of differential equations is used to describe the evolution of the power distribution network situation.

4. The distribution network situation simulation method considering multi-type source-load interaction as described in claim 1, characterized in that, The power AI model is a model with billions of parameters.

5. The distribution network situation simulation method considering multi-type source-load interaction as described in claim 4, characterized in that, The specific steps for constructing a large-scale AI model for the power industry and pre-training it using an autoregressive method are as follows: A large-scale AI model for the power industry was built based on deep learning technology and the Transformer architecture. The objective function of the autoregressive method is constructed with the goal of maximizing the conditional probability of future states given historical states. Based on the objective function, an autoencoder algorithm is used to pre-train a large-scale AI model for the power industry.

6. The distribution network situation simulation method considering multi-type source-load interaction as described in claim 5, characterized in that, The large-scale AI model for power systems includes an input embedding layer, an encoder module, a decoder module, a multi-head attention mechanism module, a feedforward neural network module, and an output projection layer.

7. The distribution network situation simulation method considering multi-type source-load interaction as described in claim 1, characterized in that, Multiple fine-tuning techniques, including Full Fine-tuning, LoRA, and Prompt-tuning, are used to perform full parameter updates, low-rank decomposition, and soft hint optimization on the large power AI model, respectively.

8. A distribution network situation simulation system considering multi-type source-load interaction, characterized in that, include: The data acquisition module is configured to acquire the operating status of various types of source loads and power distribution equipment in the power distribution network; The collaborative relationship prediction module is configured to capture the collaborative prediction relationship between source load and power distribution equipment using a multivariate time series coupling model; The model training module is configured to build a large-scale AI model for the power industry and to pre-train the large-scale AI model for the power industry using an autoregressive method. The fine-tuning module is configured to use a combination of fine-tuning techniques to fine-tune the parameters of the pre-trained power AI large model. The situation simulation module is configured to use a fine-tuned power AI model to simulate the situation of the power distribution network based on the operating status of the power sources and distribution equipment, and the collaborative prediction relationship between the power sources and distribution equipment.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-7, which is a distribution network situation simulation method considering multiple types of source-load interactions.

10. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the distribution network situation simulation method considering multi-type source-load interaction as described in any one of claims 1-7.

Images