Power distribution network abnormal risk early warning method based on multi-mode space-time large model

Abstract

The invention relates to the technical field of power distribution network risk early warning, in particular to a power distribution network abnormal risk early warning method based on a multi-mode space-time large model. According to the method, a mapping relation between multi-dimensional influence factors such as a source load space-time matching characteristic, voltage sensitivity, a net load peak value, a power distribution network topological structure and meteorological data and a new energy power distribution network abnormal risk is deeply mined based on a Lama large model; and respectively constructing independent optimization branches of power distribution network node voltage out-of-limit prediction and transformer area distribution transformer weight and overload prediction, and balancing a synchronous optimization target of prediction of different abnormal risk events through a multi-task weighted loss function. The method comprises the following steps: aligning text information and spatio-temporal information of historical marking abnormal risks of a new energy power distribution network by adopting a spatio-temporal instruction fine tuning model, dynamically adjusting model parameters, and generating an abnormal risk auxiliary prediction threshold in combination with a risk judgment threshold; therefore, occurrence moments and positions of abnormal risks such as node voltage out-of-limit of the new energy power distribution network and heavy and overload of the distribution transformer in the transformer area can be accurately predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network risk warning, and in particular to a distribution network abnormality risk warning method based on a multimodal spatiotemporal large model. Background Art

[0002] With the large-scale access of new energy equipment such as distributed photovoltaics and electric vehicle charging piles to the distribution network, the new energy distribution network presents complex operating characteristics with strong uncertainty on both the source and load sides and multi-time scale coupling; this leads to frequent abnormal risks such as node voltage exceeding the limit and heavy overload of distribution transformers in substations. In this context, how to accurately predict the time and location of occurrence of abnormal risks in the distribution network has become a key issue to ensure the safe and stable operation of the power grid.

[0003] On the one hand, predicting voltage overruns at distribution network nodes and heavy overloads at distribution transformer substations requires comprehensive consideration of multimodal data, including load time series data, voltage time series data, distribution network topology, and meteorological data. These data exhibit spatiotemporal heterogeneity and exhibit strong nonlinear interactions. Efficiently and accurately extracting key information from this complex data environment presents a core challenge in achieving accurate risk prediction. On the other hand, the spatiotemporal randomness of distributed photovoltaic output and load fluctuations leads to increased voltage fluctuations at distribution network nodes and frequent bidirectional power flows on feeders. Existing early warning models lack in-depth understanding of the multidimensional factors influencing distribution network anomaly risks, making it difficult to accurately capture dynamic risks. Current prediction methods for voltage overruns at distribution network nodes and heavy overloads at distribution transformer substations in renewable energy distribution networks primarily include statistical analysis based on historical data correlation, steady-state power flow analysis, and simulation model analysis based on physics-data fusion. However, these methods primarily rely on single-dimensional data analysis and lack the ability to comprehensively process the spatiotemporal characteristics of renewable energy distribution networks and the integration of multimodal data, making it difficult to accurately predict the time and location of anomaly risks. In addition, the above methods usually set fixed thresholds for graded warnings of single indicators, without considering the comprehensive impact of multi-dimensional indicators such as over-limit amplitude, duration, and frequency of occurrence, and lack the adaptive ability to warn of abnormal risks in distribution networks under different wiring modes; in summary, the accuracy of existing abnormal risk warning schemes for new energy distribution networks is not high. Summary of the Invention

[0004] The main purpose of the present invention is to provide a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model, aiming to solve the problem of low accuracy of existing new energy distribution network abnormality risk early warning schemes.

[0005] The technical solution proposed by the present invention is:

[0006] A distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model is applied to a distribution network abnormality risk early warning system based on a multimodal spatiotemporal large model; the system includes a data acquisition module and a calculation module that are communicatively connected to each other; the method includes:

[0007] The calculation module deeply mines the nonlinear mapping relationship between multi-dimensional influencing factors and distribution network abnormality risks based on the Llama large model, wherein the multi-dimensional influencing factors include the source-load spatiotemporal matching characteristics, voltage sensitivity, net load peak, distribution network topology, and meteorological data collected by the acquisition module;

[0008] The calculation module constructs the distribution network node voltage over-limit prediction optimization branch and the substation distribution transformer heavy overload prediction optimization branch respectively, and balances the synchronous optimization objectives of different abnormal risk event predictions through a multi-task weighted loss function;

[0009] The calculation module uses a spatiotemporal instruction fine-tuning model to align the text information and spatiotemporal information of the historical marking abnormal risks of the new energy distribution network, dynamically adjusts the parameters of the spatiotemporal instruction fine-tuning model, and combines the risk judgment threshold to generate an auxiliary prediction threshold for abnormal risks, thereby predicting the time and location of occurrence of abnormal risks in the new energy distribution network, wherein the abnormal risks of the new energy distribution network include node voltage exceeding the limit and heavy overload of distribution transformers in the substation.

[0010] Preferably, the calculation module deeply mines the nonlinear mapping relationship between multi-dimensional influencing factors and distribution network abnormality risks based on the Llama large model, including:

[0011] The calculation module normalizes the time series data of the distribution transformer load, the voltage time series data of the distribution network nodes, and the distributed photovoltaic output time series data collected by the distribution automation system and the advanced measurement system, converting the average value of each time series to 0 and the standard deviation to 1, which is expressed as:

[0012]

[0013] Where t is the sampling time, T is the length of the time series obtained by sampling, n is the number of distribution network substations, and k is the number of distribution network topology nodes; The load time series data of the distribution transformer n in the substation; is the voltage time series data of distribution network node n; is the distributed photovoltaic output power of node k at time t;

[0014] The calculation module is based on the Transformers autoregressive architecture of the Llama large model. It uses the spatiotemporal attention mechanism to explore the nonlinear mapping relationship between the spatiotemporal matching characteristics of source and load, voltage sensitivity, net load peak, distribution network topology, and meteorological data and abnormal risks of the new energy distribution network. Among them, the abnormal risks of the new energy distribution network include node voltage exceeding the limit and heavy overload of distribution transformers in the substation.

[0015] Preferably, the calculation module is based on the Transformers autoregressive architecture of the Llama large model, and uses the spatiotemporal attention mechanism to mine the spatiotemporal matching characteristics of source and load, voltage sensitivity, net load peak, distribution network topology, and the nonlinear mapping relationship between meteorological data and abnormal risks of new energy distribution networks, including:

[0016] The calculation module comprehensively considers the temporal and spatial matching degree between the distribution transformer load and the distributed photovoltaic output, and defines the temporal and spatial matching degree of the distribution network source and load as:

[0017]

[0018] Where Λ(t,k) is the temporal and spatial matching degree of the distribution network source and load; ω(t) is the time decay weight factor, which represents the timeliness of the input data; is the distributed photovoltaic output power of node k at time t; is the load power demand of node k at time t; The node voltage change rate driven by the voltage sensitivity matrix; is the rated voltage of node k; is the topological impedance weight factor; α is the distributed photovoltaic penetration rate; β is the grid strength adjustment factor; ρ(x,y) is the spatial weight function that obeys the Gaussian distribution; is the meteorological data-source-load response gradient, which is used to quantify the marginal effect of unit irradiance G and temperature E on source-load power; τ is the time window length; Ω is the distribution network spatial domain, which includes the spatial range of all load points and distributed energy; dx is the microelement size along the x direction in the distribution network spatial domain, and dy is the microelement size along the y direction in the distribution network spatial domain;

[0019] The calculation module determines the voltage sensitivity of the k-node:

[0020]

[0021] Where, is the node admittance matrix, represents the voltage change rate of node k; is the voltage sensitivity of node k;

[0022] The calculation module determines the net load peak value of the nth distribution transformer in the distribution network:

[0023]

[0024] Where, is the net load peak value of the distribution transformer in the nth substation; F n is the set of nodes under the jurisdiction of the nth distribution transformer; is the impedance weight coefficient of the k-node, represents the upstream cumulative impedance from node k to the feeder head end, P represents the total impedance of the feeder lines in the distribution transformer of the nth substation; loss (t,n) is the network loss power value of the substation under the jurisdiction of the nth substation distribution transformer.

[0025] Preferably, the calculation module determines the net load peak value of the n-th distribution transformer in the distribution network, and then further includes:

[0026] The calculation module uses graph theory to define the distribution network topology as:

[0027] X=(K,H,W)(4),

[0028] Where X is the topology of the distribution network; K = {k1, k2, ..., k k} is the set of topological nodes of the distribution network; H is the directed edge set of the distribution network, indicating the connection direction of the lines between nodes; W is the adjacency matrix of the topological nodes of the distribution network. If there is a connection relationship between the nodes, W is 1, and if there is no connection relationship, W is 0;

[0029] The calculation module normalizes the meteorological data obtained by the acquisition module from the meteorological department to ensure that the meteorological data is consistent with the sampling interval of the distribution network node voltage and the distribution transformer load time series data. The meteorological data includes temperature, humidity, wind speed and light intensity. The meteorological data after normalization is expressed as:

[0030] Ψ(t)=[R T (t),R H (t),R W (t),R L (t)] (5),

[0031] Where Ψ(t) is the meteorological data after normalization; R T (t) represents the temperature at time t; R H (t) represents the humidity at time t; R W (t) represents the wind speed at time t; R L (t) represents the light intensity at time t;

[0032] The computing module processes the nonlinear mapping relationship between multi-dimensional influencing factors and distribution network abnormal risk characteristics through a multi-head attention mechanism. The Llama large model uses a multi-layer Transformer network as a feature mapping network.

[0033] The calculation module inputs the multi-dimensional abnormal risk impact factor into the feature mapping network in combination with the abnormal risk of the distribution network history mark. The input characteristics of the abnormal risk impact factor of the Llama large model are:

[0034] The calculation module maps the input features to different model subspaces through different weight matrices and performs the same self-attention calculation in each subspace;

[0035] The calculation module combines the self-attention calculation results of each layer subspace, and obtains the correlation feature Attention (I in ), the calculation formula of the association feature is:

[0036]

[0037] Where C p is the projection layer; R d is the residual random layer; p is the probability of the residual random layer returning to 0; Output weight matrix for the multi-dimensional influence factor of self-attention; I i is the input feature vector in the i-th subspace;

[0038] The calculation module associates the feature Attention(I in ) is input into the feedforward neural network of the Llama large model, and two linear transformations are performed through the projection layer and the connection layer to obtain the prediction feature I of the abnormal risk influencing factor out =FFN(I in ), the calculation formula of the prediction feature is:

[0039]

[0040] Where, drop is the risk feature deactivation function; SG is the risk feature activation function; b i O is the bias parameter of the self-attention feature output.

[0041] Preferably, the calculation module constructs a distribution network node voltage over-limit prediction optimization branch and a substation distribution transformer heavy overload prediction optimization branch respectively, and balances the simultaneous optimization objectives of different abnormal risk event predictions through a multi-task weighted loss function, including:

[0042] The calculation module uses an adaptive weighted loss function to optimize the uncertainty of different abnormal risk prediction tasks to determine the voltage over-limit prediction loss of the distribution network node:

[0043]

[0044] Where, L V is the voltage over-limit prediction loss of the distribution network node; N is the total number of training samples; y v (j) is the actual node voltage limit label, y v (j) takes the value of 0 or 1, y v When (j) is 1, it means that the sample is a real node voltage exceeding the limit; f v (j) is the predicted probability of node voltage exceeding the limit; j is the index of the training sample;

[0045] The calculation module determines the predicted loss of heavy overload of the distribution transformer in the substation as:

[0046]

[0047] Where, L P Predict losses for heavy overload of distribution transformer in the substation area; p (j) is the real area distribution transformer overload label, the value is 0 or 1, y p When (j) is 1, it means that the sample is a real distribution transformer overload; f p (j) is the predicted probability of heavy overload of distribution transformer in the substation area;

[0048] The calculation module uses an adaptive weighted loss function to define the total prediction loss of the Llama model:

[0049]

[0050] Where, L total is the total prediction loss; σ V is the uncertainty parameter for node voltage over-limit prediction; σ P It is the uncertainty parameter for the prediction of heavy overload of distribution transformer in the substation.

[0051] Preferably, the calculation module adopts an adaptive weighted loss function to define the total prediction loss of the Llama large model, and then further includes:

[0052] In the optimization process of the two prediction branches of distribution network node voltage exceeding limit and substation distribution transformer overload, the calculation module uses the Adam optimizer to adjust the parameters of the Llama model to minimize the total prediction loss:

[0053]

[0054] Where θ is the parameter of the Llama model; η is the predicted learning rate of the Llama model, which controls the step size of parameter update; W V with b V is the voltage weight parameter of the output network; W P with b P is the power weight parameter of the output network; is the gradient of the total prediction loss;

[0055] The calculation module uses a spatiotemporal instruction fine-tuning model to align text information and spatiotemporal information of historical abnormal risk markers of the distribution network, and fuses data from different modalities to form a more accurate information representation.

[0056] Preferably, the calculation module uses a spatiotemporal instruction fine-tuning model to align text information and spatiotemporal information of the distribution network history marking abnormal risk, and fuses data of different modalities to form a more accurate information representation, and then further includes:

[0057] The calculation module uses a decoupled dual-channel attention mechanism to define two independent optimization branches for distribution network node voltage over-limit and substation distribution transformer overload. The node voltage over-limit branch predicts whether the voltage of each node in the distribution network exceeds the limit:

[0058]

[0059] Where, is the predicted voltage value of node k at the next moment; I V (t+1,k) is the node voltage limit judgment function, I V When (t+1,k) is 1, it means the voltage of the node is out of limit; Predict the upper limit value of node voltage; Predict the lower limit value for the node voltage;

[0060] The calculation module predicts whether the load in the distribution area exceeds the rated capacity of the transformer based on the heavy overload branch of the distribution transformer in the area:

[0061]

[0062] Where, To predict the load power of the nth distribution transformer in the next moment; the heavy overload of the distribution transformer in the distribution area includes forward / reverse heavy overload, I P (t+1,n) is the feeder forward / reverse heavy overload judgment function, indicating the feeder forward / reverse heavy overload; The heavy overload threshold of the distribution transformer in the nth substation.

[0063] Preferably, the calculation module uses a spatiotemporal instruction fine-tuning model to align text information and spatiotemporal information of historical abnormal risk markers of the new energy distribution network, dynamically adjusts parameters of the spatiotemporal instruction fine-tuning model, and generates an auxiliary prediction threshold for abnormal risk in combination with a risk judgment threshold, thereby predicting the time and location of occurrence of abnormal risk in the new energy distribution network, and then further includes:

[0064] The calculation module quantitatively evaluates the estimated abnormal risk of the distribution network by considering multiple indicators such as the over-limit amplitude, duration, and frequency of occurrence, including:

[0065] The calculation module can measure the degree to which the distribution network node voltage and the distribution transformer load power deviate from the rated value based on the over-limit amplitude index:

[0066]

[0067] Where, δ V is the voltage over-limit value index; δ p It is the power limit value indicator; is the rated voltage of node k; It is the rated load power of the distribution transformer in the nth substation.

[0068] Preferably, the calculation module can measure the degree to which the distribution network node voltage and the distribution transformer load power deviate from the rated value based on the over-limit amplitude index, and then further includes:

[0069] The calculation module determines the duration T of the mth independent abnormal risk event in the new energy distribution network. m is the time span from the start to the end of the event, as shown in the following formula:

[0070]

[0071] Where, T m is the time span; is the starting time of the mth abnormal risk event; is the end time of the mth abnormal risk event;

[0072] The calculation module determines the frequency F of the mth independent abnormal risk event in the new energy distribution network. m is the average number of occurrences of the event per unit time, as shown in the following formula:

[0073]

[0074] Where N m is the total number of occurrences of the mth abnormal risk event within the statistical period ΔT.

[0075] Preferably, the calculation module quantitatively evaluates the estimated abnormal risk of the distribution network by considering multi-dimensional indicators such as over-limit amplitude, duration and occurrence frequency, and then further includes:

[0076] The calculation module uses a multi-level warning mechanism combined with warning levels to provide graded warnings for the estimated abnormal risks of the new energy distribution network, including:

[0077] The calculation module uses q-order orthopair hesitant fuzzy to effectively integrate multi-dimensional evaluation indicators and construct an abnormal risk warning fuzzy decision matrix:

[0078]

[0079] Where μ ma is the membership degree, which indicates the degree to which the indicator a belongs to the better solution in the mth abnormal risk event. When a = 1, it corresponds to the frequency of occurrence; when a = 2, it corresponds to the limit-crossing value; when a = 3, it corresponds to the duration; v ma is the non-membership degree, which indicates the degree to which indicator a does not belong to the optimal solution in the mth abnormal risk event; w a is the weight of indicator a; q is the fuzzy matrix parameter; M is the total number of abnormal risk events;

[0080] The calculation module takes the best / worst membership combination for each indicator to obtain the positive / negative ideal solution of the multi-dimensional abnormal risk warning indicator:

[0081]

[0082] Where, is the positive ideal solution of the abnormal risk warning indicator; is the negative ideal solution of the abnormal risk warning indicator; cost-type indicators such as over-limit amplitude and duration take the minimum membership, and the inverse of the occurrence frequency indicator takes the maximum membership;

[0083] The calculation module uses the improved weighted q-order Hamming distance to calculate the similarity between the m-th abnormal risk event and the positive ideal solution to obtain the warning degree C m :

[0084]

[0085] Where, is the mth abnormal risk event and the positive ideal solution distance, The larger it is, the higher the risk of the mth abnormal risk event; is the mth abnormal risk event and the negative ideal solution h j - distance, The larger the value, the lower the risk of the mth abnormal risk event;ma is the decision element of the abnormal risk warning fuzzy decision matrix; C m is the warning degree of the mth abnormal risk event, C m The closer it is to 1, the lower the warning level of the mth abnormal risk event is. m The closer it is to 0, the higher the warning level of the mth abnormal risk event.

[0086] The above technical solution can achieve the following beneficial effects:

[0087] The distribution network abnormal risk warning method based on a multimodal spatiotemporal large model proposed in the present invention deeply mines the mapping relationship between multi-dimensional influencing factors such as source-load spatiotemporal matching characteristics, voltage sensitivity, net load peak, distribution network topology, and meteorological data and abnormal risks of new energy distribution networks based on the Llama large model; constructs independent optimization branches for distribution network node voltage over-limit prediction and substation distribution transformer overload prediction respectively, and balances the simultaneous optimization goals of different abnormal risk event predictions through a multi-task weighted loss function. A spatiotemporal instruction fine-tuning model is used to align the text information and spatiotemporal information of the historical abnormal risk marking of the new energy distribution network, dynamically adjust the model parameters and generate an auxiliary prediction threshold for abnormal risk in combination with the risk judgment threshold, thereby accurately predicting the time and location of abnormal risks such as voltage over-limit of new energy distribution network nodes and heavy overload of substation distribution transformers. Compared with the existing new energy distribution network abnormal risk warning scheme, its accuracy is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0089] Figure 1 This is a flowchart of the first embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention;

[0090] Figure 2 This is a topological diagram of an IEEE 33-node distribution network according to the tenth embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention;

[0091] Figure 3 This is a comparison chart of the distribution network abnormality risk warning accuracy of various algorithms of the tenth embodiment of the distribution network abnormality risk warning method based on a multimodal spatiotemporal large model proposed by the present invention. DETAILED DESCRIPTION

[0092] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0093] The present invention proposes a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model.

[0094] As attached Figure 1 As shown, in a first embodiment of a distribution network abnormality risk warning method based on a multimodal spatiotemporal large model proposed by the present invention, the method is applied to a distribution network abnormality risk warning system based on a multimodal spatiotemporal large model; the system includes a data acquisition module and a calculation module that are communicatively connected to each other; this embodiment includes the following steps:

[0095] Step S110: The calculation module deeply mines the nonlinear mapping relationship between multidimensional influencing factors and distribution network abnormality risks based on the Llama large model, wherein the multidimensional influencing factors include the source-load time-space matching characteristics, voltage sensitivity, net load peak, distribution network topology, and meteorological data collected by the acquisition module.

[0096] Specifically, the Llama model is a series of large language models developed by Meta, which have undergone multiple iterations and upgrades. Llama models are known for their open source nature and multimodal processing capabilities, enabling content conversion between multiple data types such as text, video, images, and audio.

[0097] Step S120: The calculation module constructs a distribution network node voltage over-limit prediction optimization branch and a substation distribution transformer heavy overload prediction optimization branch respectively, and balances the synchronous optimization objectives of different abnormal risk event predictions through a multi-task weighted loss function.

[0098] Step S130: The calculation module uses the spatiotemporal instruction fine-tuning model to align the text information and spatiotemporal information of the historical marking abnormal risks of the new energy distribution network, dynamically adjusts the parameters of the spatiotemporal instruction fine-tuning model, and combines the risk judgment threshold to generate an auxiliary prediction threshold for abnormal risks, thereby predicting the time and location of occurrence of abnormal risks in the new energy distribution network, wherein the abnormal risks of the new energy distribution network include node voltage exceeding the limit and heavy overload of distribution transformers in the substation.

[0099] The distribution network abnormal risk warning method based on a multimodal spatiotemporal large model proposed in the present invention deeply mines the mapping relationship between multi-dimensional influencing factors such as source-load spatiotemporal matching characteristics, voltage sensitivity, net load peak, distribution network topology, and meteorological data and abnormal risks of new energy distribution networks based on the Llama large model; constructs independent optimization branches for distribution network node voltage over-limit prediction and substation distribution transformer overload prediction respectively, and balances the simultaneous optimization goals of different abnormal risk event predictions through a multi-task weighted loss function. A spatiotemporal instruction fine-tuning model is used to align the text information and spatiotemporal information of the historical abnormal risk marking of the new energy distribution network, dynamically adjust the model parameters and generate an auxiliary prediction threshold for abnormal risk in combination with the risk judgment threshold, thereby accurately predicting the time and location of abnormal risks such as voltage over-limit of new energy distribution network nodes and heavy overload of substation distribution transformers. Compared with the existing new energy distribution network abnormal risk warning scheme, its accuracy is higher.

[0100] In a second embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the first embodiment, step S110 includes the following steps:

[0101] Step S210: The calculation module normalizes the substation distribution transformer load time series data, distribution network node voltage time series data, and distributed photovoltaic output time series data collected by the distribution automation system and the advanced measurement system, converting the average value of each time series to 0 and the standard deviation to 1, expressed as:

[0102]

[0103] Where t is the sampling time, T is the length of the time series obtained by sampling, n is the number of distribution network substations, and k is the number of distribution network topology nodes; The load time series data of the distribution transformer n in the substation; is the voltage time series data of distribution network node n; is the distributed photovoltaic output power of node k at time t.

[0104] Step S220: The calculation module is based on the Transformers autoregressive architecture of the Llama large model, and uses the spatiotemporal attention mechanism to mine the nonlinear mapping relationship between source-load spatiotemporal matching characteristics, voltage sensitivity, net load peak, distribution network topology, and meteorological data and abnormal risks of the new energy distribution network, wherein the abnormal risks of the new energy distribution network include node voltage exceeding the limit and heavy overload of distribution transformers in the substation.

[0105] In a third embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the second embodiment, step S220 includes the following steps:

[0106] Step S310: The calculation module comprehensively considers the temporal and spatial matching degree between the distribution transformer load and the distributed photovoltaic output in the substation area, and defines the temporal and spatial matching degree of the distribution network source and load as:

[0107]

[0108] Where Λ(t,k) is the temporal and spatial matching degree of the distribution network source and load; ω(t) is the time decay weight factor, which represents the timeliness of the input data; is the distributed photovoltaic output power of node k at time t; is the load power demand of node k at time t; The node voltage change rate driven by the voltage sensitivity matrix; is the rated voltage of node k; is the topological impedance weight factor; α is the distributed photovoltaic penetration rate; β is the grid strength adjustment factor; ρ(x,y) is the spatial weight function that obeys the Gaussian distribution; is the meteorological data-source-load response gradient, which is used to quantify the marginal effect of unit irradiance G and temperature E on the source-load power; τ is the time window length; Ω is the distribution network spatial domain, which includes the spatial range of all load points and distributed energy; dx is the microelement size along the x direction in the distribution network spatial domain, and dy is the microelement size along the y direction in the distribution network spatial domain.

[0109] Step S320: The calculation module determines the voltage sensitivity of the k node:

[0110]

[0111] Where, is the node admittance matrix, represents the voltage change rate of node k; is the voltage sensitivity of node k.

[0112] Step S330: The calculation module determines the net load peak value of the nth distribution transformer in the distribution network:

[0113]

[0114] Where, is the net load peak value of the distribution transformer in the nth substation; F n is the set of nodes under the jurisdiction of the nth distribution transformer; is the impedance weight coefficient of the k-node, represents the upstream cumulative impedance from node k to the feeder head end, P represents the total impedance of the feeder lines in the distribution transformer of the nth substation; loss (t,n) is the network loss power value of the substation under the jurisdiction of the nth substation distribution transformer.

[0115] In a fourth embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the third embodiment, step S330 further includes the following steps:

[0116] Step S410: The calculation module uses graph theory to define the distribution network topology as:

[0117] X=(K,H,W)(4),

[0118] Where X is the topology of the distribution network; K = {k1, k2, ..., k k} is the set of distribution network topology nodes; H is the directed edge set of the distribution network, indicating the connection direction of the lines between nodes; W is the adjacency matrix of the distribution network topology nodes. If there is a connection relationship between the nodes, W is 1, and if there is no connection relationship, W is 0.

[0119] Step S420: The calculation module normalizes the meteorological data obtained by the acquisition module from the meteorological department to ensure that the meteorological data is consistent with the sampling interval of the distribution network node voltage and the distribution transformer load time series data. The meteorological data includes temperature, humidity, wind speed and light intensity. The meteorological data after normalization is expressed as:

[0120] Ψ(t)=[R T (t),R H (t),R W (t),R L (t)] (5),

[0121] Where Ψ(t) is the meteorological data after normalization; R T (t) represents the temperature at time t; R H (t) represents the humidity at time t; R W (t) represents the wind speed at time t; R L (t) represents the light intensity at time t.

[0122] Step S430: The calculation module processes the nonlinear mapping relationship between multi-dimensional influencing factors and distribution network abnormal risk characteristics through a multi-head attention mechanism, wherein the Llama large model adopts a multi-layer Transformer network as a feature mapping network.

[0123] Step S440: The calculation module inputs the multi-dimensional abnormal risk impact factor into the feature mapping network in combination with the abnormal risk of the distribution network history mark. The input feature of the abnormal risk impact factor of the Llama large model is

[0124] Step S450: The calculation module maps the input features to different model subspaces through different weight matrices, and performs the same self-attention calculation in each subspace.

[0125] Step S460: The calculation module combines the self-attention calculation results of each layer subspace, and obtains the correlation feature Attention (I in ), the calculation formula of the association feature is:

[0126]

[0127] Where C p is the projection layer; R d is the residual random layer; p is the probability of the residual random layer returning to 0; Output weight matrix for the multi-dimensional influence factor of self-attention; I i is the input feature vector in the i-th subspace.

[0128] Step S470: The calculation module associates the feature Attention(I in ) is input into the feedforward neural network (FFN) of the Llama large model, and two linear transformations are performed through the projection layer and the connection layer to obtain the prediction feature I of the abnormal risk influencing factor out =FFN(I in ), the calculation formula of the prediction feature is:

[0129]

[0130] Where, drop is the risk feature deactivation function; SG is the risk feature activation function; is the bias parameter of the self-attention feature output.

[0131] In a fifth embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the fourth embodiment, step S120 includes the following steps:

[0132] Step S510: To resolve the conflict and imbalance issues that may arise when the model simultaneously predicts node voltage over-limit and forward / reverse heavy overload of the distribution transformer in the substation, and to improve the dynamic adaptability of the model, the calculation module uses an adaptive weighted loss function to optimize the uncertainty of different abnormal risk prediction tasks to determine the distribution network node voltage over-limit prediction loss:

[0133]

[0134] Where, L Vis the voltage over-limit prediction loss of the distribution network node; N is the total number of training samples; y v (j) is the actual node voltage limit label, y v (j) takes the value of 0 or 1, y v When (j) is 1, it means that the sample is a real node voltage exceeding the limit; f v (j) is the predicted probability of node voltage exceeding the limit; j is the index of the training sample.

[0135] Step S520: The calculation module determines the predicted loss of heavy overload of the distribution transformer in the substation as:

[0136]

[0137] Where, L P Predict losses for heavy overload of distribution transformer in the substation area; p (j) is the real area distribution transformer overload label, the value is 0 or 1, y p When (j) is 1, it means that the sample is a real distribution transformer overload; f p (j) is the predicted probability of heavy overload of the distribution transformer in the substation.

[0138] Step S530: The calculation module uses an adaptive weighted loss function to define the total prediction loss of the Llama large model:

[0139]

[0140] Where, L total is the total prediction loss; σ V is the uncertainty parameter for node voltage over-limit prediction; σ P It is the uncertainty parameter for the prediction of heavy overload of distribution transformer in the substation.

[0141] In a sixth embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the fifth embodiment, step S530 further includes the following steps:

[0142] Step S610: The calculation module optimizes the parameters of the Llama model (including W V ,b V ,W P ,b P ,σ V ,σ P ) is tuned using the Adam optimizer to minimize the total prediction loss:

[0143]

[0144] Where θ is the parameter of the Llama model; η is the predicted learning rate of the Llama model, which controls the step size of parameter update; W V with b V is the voltage weight parameter of the output network; W P with b P is the power weight parameter of the output network; is the gradient of the total prediction loss.

[0145] Step S620: The calculation module uses the spatiotemporal instruction fine-tuning model to align the text information and spatiotemporal information of the distribution network history marking abnormal risks, and fuses data of different modalities to form a more accurate information representation.

[0146] Specifically, by integrating the distribution network abnormal risk features from text and spatiotemporal domains, the complementary information of spatiotemporal data is captured and extracted into a more expressive high-level semantic representation, and the historical abnormal risk spatiotemporal features are combined using a lightweight alignment module. and the spatiotemporal characteristics of multidimensional impact factors Projected into the hidden layer of the model, it is represented in the model as a special symbol of instruction.

[0147] For example: Where:<ST_start> is the starting space-time token;<ST_end> To terminate the space-time token; Marking spatiotemporal tokens for abnormal risks; It is a spatiotemporal token of multidimensional impact factors.

[0148] The parameters of the Llama large model are fine-tuned using spatiotemporal tokens to generate auxiliary prediction thresholds, and output abnormal risk prediction results such as voltage exceeding the limit at the node of the new energy distribution network and heavy overload of the distribution transformer in the substation.

[0149] In a seventh embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the sixth embodiment, step S620 further includes the following steps:

[0150] Step S710: The calculation module adopts a decoupled dual-channel attention mechanism in the prediction layer architecture design of the model to define two independent optimization branches: the voltage exceeding the limit of the distribution network node and the heavy overload of the distribution transformer in the substation. The node voltage exceeding limit branch predicts whether the voltage of each node in the distribution network exceeds the limit:

[0151]

[0152] Where, is the predicted voltage value of node k at the next moment; I V (t+1,k) is the node voltage limit judgment function, I VWhen (t+1,k) is 1, it means the voltage of the node is out of limit; Predict the upper limit value of node voltage; Predict the lower limit value for the node voltage.

[0153] Step S720: The calculation module predicts whether the load in the distribution area exceeds the rated capacity of the transformer based on the heavy overload branch of the distribution transformer in the area:

[0154]

[0155] Where, To predict the load power of the nth distribution transformer in the next moment; the heavy overload of the distribution transformer in the distribution area includes forward / reverse heavy overload, I P (t+1,n) is the feeder forward / reverse heavy overload judgment function, indicating the feeder forward / reverse heavy overload; The heavy overload threshold of the distribution transformer in the nth substation.

[0156] In an eighth embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the seventh embodiment, step S130 further includes the following steps:

[0157] Step S810: The calculation module quantitatively evaluates the estimated abnormal risk of the distribution network by considering multiple dimensional indicators such as the over-limit value, duration, and occurrence frequency, including the following steps:

[0158] Step S811: The calculation module can measure the degree to which the voltage at the distribution network node and the load power of the distribution transformer in the substation deviate from the rated value based on the over-limit value index:

[0159]

[0160] Where, δ V is the voltage over-limit value index; δ p It is the power limit value indicator; is the rated voltage of node k; It is the rated load power of the distribution transformer in the nth substation.

[0161] In a ninth embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the eighth embodiment, step S811 further includes the following steps:

[0162] Step S910: The calculation module determines the duration T of the mth independent abnormal risk event in the new energy distribution network. m is the time span from the start to the end of the event, as shown in the following formula:

[0163]

[0164] Where, T m is the time span; is the starting time of the mth abnormal risk event; is the end time of the mth abnormal risk event.

[0165] Step S920: The calculation module determines the frequency F of the mth independent abnormal risk event in the new energy distribution network. m is the average number of occurrences of the event per unit time, as shown in the following formula:

[0166]

[0167] Where N m is the total number of occurrences of the mth abnormal risk event within the statistical period ΔT.

[0168] Specifically, under different distribution network connection modes, the system voltage stability and reliability vary significantly. Therefore, when constructing an abnormal risk warning fuzzy decision matrix that integrates benefit-based indicators (occurrence frequency) and cost-based indicators (overshoot amplitude, duration), the weights of each indicator need to be set differently. For a single-radial connection distribution network, due to its single power supply, the system voltage stability is poor and the fault recovery period is long, so the overshoot amplitude and duration are assigned higher weights. In a single-ring connection distribution network, the load balancing capability brought by the closed ring operation reduces the weight of the overshoot amplitude indicator, but the frequent operation of the tie switch increases the weight of the occurrence frequency indicator. The multi-supply and one-backup connection distribution network significantly improves voltage quality due to the support of multiple power sources, so its overshoot amplitude indicator has the lowest weight, but the transient impact during the backup power supply switching process leads to a higher weight of the occurrence frequency indicator. The multi-segment and multi-tie connection distribution network dynamically reconfigures the network topology through the sectionalizer, which reduces the weight of the duration indicator to the lowest. However, the power fluctuations caused by the high proportion of new energy access significantly increase the frequency of operating mode switching, so the occurrence frequency indicator has the highest weight. The weights assigned to each indicator under different distribution network connection modes are shown in Table 1.

[0169] Table 1: Weights assigned to various indicators under different connection modes of distribution network

[0170]

[0171] In a tenth embodiment of a distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model proposed by the present invention, based on the eighth embodiment, step S810 further includes the following steps:

[0172] Step S1010: The calculation module uses a multi-level warning mechanism combined with a warning degree to perform a graded warning on the estimated abnormal risk of the new energy distribution network, including the following steps:

[0173] Step S1020: The calculation module uses the q-order orthopair hesitant fuzzy (TOPSIS method) to effectively integrate multi-dimensional evaluation indicators and construct an abnormal risk warning fuzzy decision matrix:

[0174]

[0175] Where μ ma is the membership degree, which indicates the degree to which the indicator a belongs to the better solution in the mth abnormal risk event. When a = 1, it corresponds to the frequency of occurrence; when a = 2, it corresponds to the limit-crossing value; when a = 3, it corresponds to the duration; v ma is the non-membership degree, which indicates the degree to which indicator a does not belong to the optimal solution in the mth abnormal risk event; w a is the weight of indicator a; q is the fuzzy matrix parameter; M is the total number of abnormal risk events.

[0176] Step S1030: The calculation module takes the best / worst membership combination for each indicator to obtain the positive / negative ideal solution of the multi-dimensional abnormal risk warning indicator:

[0177]

[0178] Where, is the positive ideal solution of the abnormal risk warning indicator; is the negative ideal solution of the abnormal risk warning indicator; cost-type indicators such as the over-limit amplitude and duration take the minimum membership, and the inverse of the occurrence frequency indicator takes the maximum membership.

[0179] Step S1040: The calculation module uses the improved weighted q-order Hamming distance (Hamming distance) to calculate the similarity between the m-th abnormal risk event and the positive ideal solution to obtain the warning degree C m :

[0180]

[0181] Where, is the mth abnormal risk event and the positive ideal solution distance, The larger it is, the higher the risk of the mth abnormal risk event; is the mth abnormal risk event and the negative ideal solution distance, The larger the value, the lower the risk of the mth abnormal risk event; ma is the decision element of the abnormal risk warning fuzzy decision matrix; C m is the warning degree of the mth abnormal risk event, C m The closer it is to 1, the lower the warning level of the mth abnormal risk event is. mThe closer it is to 0, the higher the warning level of the mth abnormal risk event. The classification criteria for the warning level of abnormal risk events are shown in Table 2.

[0182] Table 2: New energy distribution network abnormal risk warning level classification standards

[0183]

[0184]

[0185] Case analysis:

[0186] The model and method proposed in this scheme are applied to the improved IEEE33-node distribution network. Figure 2 As shown, PV systems are connected at nodes 12, 22, and 30, and four electric vehicle charging stations are connected at nodes 8, 21, 27, and 32. The main transformer capacity is 5MVA, and the rated capacity of the distribution substations is 400kVA. The total installed PV capacity reaches 40% of the peak load of the distribution network.

[0187] The results of the abnormal risk warning of the distribution network on a certain day are shown in Table 3. The comparison of the accuracy of the abnormal risk warning of the distribution network using various algorithms is shown in the attached figure. Figure 3 As shown, four abnormal risk warning methods are considered for comparison: Method 1: the method proposed in this patent; Method 2: the abnormal risk warning method based on statistical analysis; Method 3: the abnormal risk warning method based on steady-state power flow analysis; Method 4: the abnormal risk warning method based on simulation analysis.

[0188] The results show that the accuracy and recall rate of abnormal risk warning of the method proposed in this patent are higher than those of other commonly used algorithms. The recognition accuracy rate can be improved by at least 6.88%, and the recognition recall rate can be improved by at least 6.37%. The calculation formulas for the recognition accuracy rate and the recognition recall rate are shown in formula (22).

[0189]

[0190] Where F1 represents the detection accuracy; F2 represents the detection recall; TP is the number of samples correctly detected as abnormal risk events; TN is the number of samples correctly detected as normal events; FP is the number of samples of normal events that are incorrectly detected as abnormal risk events; and FN is the number of samples of abnormal risk events that are incorrectly detected as normal events.

[0191] Table 3: Early warning results of abnormal events in the distribution network on a certain day

[0192]

[0193] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0194] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A distribution network abnormality risk warning method based on a multimodal spatiotemporal large model, characterized by: The invention is applied to a distribution network abnormality risk early warning system based on a multimodal spatiotemporal large model; the system includes a data acquisition module and a calculation module that are communicatively connected to each other; the method includes: The calculation module deeply mines the nonlinear mapping relationship between multi-dimensional influencing factors and distribution network abnormality risks based on the Llama large model, wherein the multi-dimensional influencing factors include the source-load spatiotemporal matching characteristics, voltage sensitivity, net load peak, distribution network topology, and meteorological data collected by the acquisition module; The calculation module constructs the distribution network node voltage over-limit prediction optimization branch and the substation distribution transformer heavy overload prediction optimization branch respectively, and balances the synchronous optimization objectives of different abnormal risk event predictions through a multi-task weighted loss function; The calculation module uses a spatiotemporal instruction fine-tuning model to align the text information and spatiotemporal information of the historical marking abnormal risks of the new energy distribution network, dynamically adjusts the parameters of the spatiotemporal instruction fine-tuning model, and combines the risk judgment threshold to generate an auxiliary prediction threshold for abnormal risks, thereby predicting the time and location of occurrence of abnormal risks in the new energy distribution network, wherein the abnormal risks of the new energy distribution network include node voltage exceeding the limit and heavy overload of distribution transformers in the substation.

2. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 1 is characterized in that: The calculation module deeply mines the nonlinear mapping relationship between multi-dimensional influencing factors and distribution network abnormality risks based on the Llama large model, including: The calculation module normalizes the time series data of the distribution transformer load, the voltage time series data of the distribution network nodes, and the distributed photovoltaic output time series data collected by the distribution automation system and the advanced measurement system, converting the average value of each time series to 0 and the standard deviation to 1, which is expressed as: Where t is the sampling time, T is the length of the time series obtained by sampling, n is the number of distribution network substations, and k is the number of distribution network topology nodes; is the load time series data of the distribution transformer n in the substation; V t k is the voltage time series data of distribution network node n; is the distributed photovoltaic output power of node k at time t; The calculation module is based on the Transformers autoregressive architecture of the Llama large model. It uses the spatiotemporal attention mechanism to explore the nonlinear mapping relationship between the spatiotemporal matching characteristics of source and load, voltage sensitivity, net load peak, distribution network topology, and meteorological data and abnormal risks of the new energy distribution network. Among them, the abnormal risks of the new energy distribution network include node voltage exceeding the limit and heavy overload of distribution transformers in the substation.

3. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 2 is characterized in that: The calculation module is based on the Transformers autoregressive architecture of the Llama large model. It uses the spatiotemporal attention mechanism to explore the spatiotemporal matching characteristics of source and load, voltage sensitivity, net load peak, distribution network topology, and the nonlinear mapping relationship between meteorological data and abnormal risks of new energy distribution networks, including: The calculation module comprehensively considers the temporal and spatial matching degree between the distribution transformer load and the distributed photovoltaic output, and defines the temporal and spatial matching degree of the distribution network source and load as: Where Λ(t,k) is the temporal and spatial matching degree of the distribution network source and load; ω(t) is the time decay weight factor, which represents the timeliness of the input data; is the distributed photovoltaic output power of node k at time t; is the load power demand of node k at time t; The node voltage change rate driven by the voltage sensitivity matrix; is the rated voltage of node k; is the topological impedance weight factor; α is the distributed photovoltaic penetration rate; β is the grid strength adjustment factor; ρ(x,y) is the spatial weight function that obeys the Gaussian distribution; is the meteorological data-source-load response gradient, which is used to quantify the marginal effect of unit irradiance G and temperature E on source-load power; τ is the time window length; Ω is the distribution network spatial domain, which includes the spatial range of all load points and distributed energy; dx is the microelement size along the x direction in the distribution network spatial domain, and dy is the microelement size along the y direction in the distribution network spatial domain; The calculation module determines the voltage sensitivity of the k-node: Where, is the node admittance matrix, represents the voltage change rate of node k; is the voltage sensitivity of node k; The calculation module determines the net load peak value of the nth distribution transformer in the distribution network: Where, is the net load peak value of the distribution transformer in the nth substation; F n is the set of nodes under the jurisdiction of the nth distribution transformer; is the impedance weight coefficient of the k-node, represents the upstream cumulative impedance from node k to the feeder head end, P represents the total impedance of the feeder lines in the distribution transformer of the nth substation; loss (t,n) is the network loss power value of the substation under the jurisdiction of the nth substation distribution transformer.

4. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 3 is characterized in that: The calculation module determines the net load peak of the n-th distribution transformer in the distribution network, and then further includes: The calculation module uses graph theory to define the distribution network topology as: X=(K,H,W) (4), Where X is the topology of the distribution network; K = {k1, k2, ..., k k } is the set of topological nodes of the distribution network; H is the directed edge set of the distribution network, indicating the connection direction of the lines between nodes; W is the adjacency matrix of the topological nodes of the distribution network. If there is a connection relationship between the nodes, W is 1, and if there is no connection relationship, W is 0; The calculation module normalizes the meteorological data obtained by the acquisition module from the meteorological department to ensure that the meteorological data is consistent with the sampling interval of the distribution network node voltage and the distribution transformer load time series data. The meteorological data includes temperature, humidity, wind speed and light intensity. The meteorological data after normalization is expressed as: Ψ(t)=[R T (t),R H (t),R W (t),R L (t)] (5), Where Ψ(t) is the meteorological data after normalization; R T (t) represents the temperature at time t; R H (t) represents the humidity at time t; R W (t) represents the wind speed at time t; R L (t) represents the light intensity at time t; The computing module processes the nonlinear mapping relationship between multi-dimensional influencing factors and distribution network abnormal risk characteristics through a multi-head attention mechanism. The Llama large model uses a multi-layer Transformer network as a feature mapping network. The calculation module inputs the multi-dimensional abnormal risk impact factor into the feature mapping network in combination with the abnormal risk of the distribution network history mark. The input characteristics of the abnormal risk impact factor of the Llama large model are: The calculation module maps the input features to different model subspaces through different weight matrices and performs the same self-attention calculation in each subspace; The calculation module combines the self-attention calculation results of each layer subspace, and obtains the correlation feature Attention (I in ), the calculation formula of the association feature is: Where C p is the projection layer; R d is the residual random layer; p is the probability of the residual random layer returning to 0; W i O Output weight matrix for the multi-dimensional influence factor of self-attention; I i is the input feature vector in the i-th subspace; The calculation module associates the feature Attention(I in ) is input into the feedforward neural network of the Llama large model, and two linear transformations are performed through the projection layer and the connection layer to obtain the prediction feature I of the abnormal risk influencing factor out =FFN(I in ), the calculation formula of the prediction feature is: Where, drop is the risk feature deactivation function; SG is the risk feature activation function; is the bias parameter of the self-attention feature output.

5. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 4 is characterized in that: The calculation module constructs the distribution network node voltage over-limit prediction optimization branch and the substation distribution transformer heavy overload prediction optimization branch respectively, and balances the simultaneous optimization objectives of different abnormal risk event predictions through a multi-task weighted loss function, including: The calculation module uses an adaptive weighted loss function to optimize the uncertainty of different abnormal risk prediction tasks to determine the voltage over-limit prediction loss of the distribution network node: Where, L V is the voltage over-limit prediction loss of the distribution network node; N is the total number of training samples; y v (j) is the actual node voltage limit label, y v (j) takes the value of 0 or 1, y v When (j) is 1, it means that the sample is a real node voltage exceeding the limit; f v (j) is the predicted probability of node voltage exceeding the limit; j is the index of the training sample; The calculation module determines the predicted loss of heavy overload of the distribution transformer in the substation as: Where, L P Predict losses for heavy overload of distribution transformer in the substation area; p (j) is the real area distribution transformer overload label, the value is 0 or 1, y p When (j) is 1, it means that the sample is a real distribution transformer overload; f p (j) is the predicted probability of heavy overload of distribution transformer in the substation area; The calculation module uses an adaptive weighted loss function to define the total prediction loss of the Llama model: Where, L total is the total prediction loss; σ V is the uncertainty parameter for node voltage over-limit prediction; σ P It is the uncertainty parameter for the prediction of heavy overload of distribution transformer in the substation.

6. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 5 is characterized in that: The calculation module uses an adaptive weighted loss function to define the total prediction loss of the Llama large model, and then also includes: In the optimization process of the two prediction branches of distribution network node voltage exceeding limit and substation distribution transformer overload, the calculation module uses the Adam optimizer to adjust the parameters of the Llama model to minimize the total prediction loss: Where θ is the parameter of the Llama model; η is the predicted learning rate of the Llama model, which controls the step size of parameter update; W V with b V is the voltage weight parameter of the output network; W P with b P is the power weight parameter of the output network; is the gradient of the total prediction loss; The calculation module uses a spatiotemporal instruction fine-tuning model to align text information and spatiotemporal information of historical abnormal risk markers of the distribution network, and fuses data from different modalities to form a more accurate information representation.

7. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 6 is characterized in that: The calculation module uses a spatiotemporal instruction fine-tuning model to align text information and spatiotemporal information of the distribution network's historical abnormal risk markers, and fuses data from different modalities to form a more accurate information representation. It also includes: The calculation module uses a decoupled dual-channel attention mechanism to define two independent optimization branches for distribution network node voltage over-limit and substation distribution transformer overload. The node voltage over-limit branch predicts whether the voltage of each node in the distribution network exceeds the limit: Where, is the predicted voltage value of node k at the next moment; I V (t+1,k) is the node voltage limit judgment function, I V When (t+1,k) is 1, it means the voltage of the node is out of limit; Predict the upper limit value of node voltage; Predict the lower limit value for the node voltage; The calculation module predicts whether the load in the distribution area exceeds the rated capacity of the transformer based on the heavy overload branch of the distribution transformer in the area: Where, To predict the load power of the nth distribution transformer in the next moment; the heavy overload of the distribution transformer in the distribution area includes forward / reverse heavy overload, I P (t+1,n) is the feeder forward / reverse heavy overload judgment function, indicating the feeder forward / reverse heavy overload; The heavy overload threshold of the distribution transformer in the nth substation.

8. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 7 is characterized in that: The calculation module uses a spatiotemporal instruction fine-tuning model to align text information and spatiotemporal information of historical abnormal risk markers of the new energy distribution network, dynamically adjusts the parameters of the spatiotemporal instruction fine-tuning model, and generates an auxiliary prediction threshold for abnormal risk in combination with a risk judgment threshold, thereby predicting the time and location of occurrence of abnormal risk in the new energy distribution network. The following also includes: The calculation module quantitatively evaluates the estimated abnormal risk of the distribution network by considering multiple indicators such as the over-limit amplitude, duration, and frequency of occurrence, including: The calculation module can measure the degree to which the distribution network node voltage and the distribution transformer load power deviate from the rated value based on the over-limit amplitude index: Where, δ V is the voltage over-limit value index; δ p It is the power limit value indicator; is the rated voltage of node k; It is the rated load power of the distribution transformer in the nth substation.

9. The method for warning abnormal risk of distribution network based on multimodal spatiotemporal large model according to claim 8 is characterized in that: The calculation module can measure the degree to which the voltage of the distribution network node and the load power of the distribution transformer in the substation deviate from the rated value based on the over-limit amplitude index, and then further includes: The calculation module determines the duration T of the mth independent abnormal risk event in the new energy distribution network. m is the time span from the start to the end of the event, as shown in the following formula: Where, T m is the time span; is the starting time of the mth abnormal risk event; is the end time of the mth abnormal risk event; The calculation module determines the frequency F of the mth independent abnormal risk event in the new energy distribution network. m is the average number of occurrences of the event per unit time, as shown in the following formula: Where N m is the total number of occurrences of the mth abnormal risk event within the statistical period ΔT.

10. The distribution network abnormality risk early warning method based on a multimodal spatiotemporal large model according to claim 8 is characterized in that: The calculation module quantitatively evaluates the estimated abnormal risk of the distribution network by considering multiple indicators such as the over-limit amplitude, duration, and frequency of occurrence, and then further includes: The calculation module uses a multi-level warning mechanism combined with warning levels to provide graded warnings for the estimated abnormal risks of the new energy distribution network, including: The calculation module uses q-order orthopair hesitant fuzzy to effectively integrate multi-dimensional evaluation indicators and construct an abnormal risk warning fuzzy decision matrix: Where μ ma is the membership degree, which indicates the degree to which the indicator a belongs to the better solution in the mth abnormal risk event. When a = 1, it corresponds to the frequency of occurrence; when a = 2, it corresponds to the limit-crossing value; when a = 3, it corresponds to the duration; v ma is the non-membership degree, which indicates the degree to which indicator a does not belong to the optimal solution in the mth abnormal risk event; w a is the weight of indicator a; q is the fuzzy matrix parameter; M is the total number of abnormal risk events; The calculation module takes the best / worst membership combination for each indicator to obtain the positive / negative ideal solution of the multi-dimensional abnormal risk warning indicator: Where, is the positive ideal solution of the abnormal risk warning indicator; is the negative ideal solution of the abnormal risk warning indicator; cost-type indicators such as over-limit amplitude and duration take the minimum membership, and the inverse of the occurrence frequency indicator takes the maximum membership; The calculation module uses the improved weighted q-order Hamming distance to calculate the similarity between the m-th abnormal risk event and the positive ideal solution to obtain the warning degree C m : Where, is the mth abnormal risk event and the positive ideal solution distance, The larger it is, the higher the risk of the mth abnormal risk event; is the mth abnormal risk event and the negative ideal solution distance, The larger the value, the lower the risk of the mth abnormal risk event; ma is the decision element of the abnormal risk warning fuzzy decision matrix; C m is the warning degree of the mth abnormal risk event, C m The closer it is to 1, the lower the warning level of the mth abnormal risk event is. m The closer it is to 0, the higher the warning level of the mth abnormal risk event.

Images