Real-time line loss analysis method and system based on cross-modal measurement data fusion

Abstract

The application relates to a real-time line loss analysis method and system based on cross-modal measurement data fusion, which comprises the following steps: S1: collecting multi-source heterogeneous data, and adopting an improved isolation forest algorithm to perform preliminary quality screening on the collected original data, obtaining an original data set and a corresponding data quality preliminary evaluation label; S2: according to the data quality preliminary evaluation label, the original data set is pretreated to obtain pretreated data; S3: according to the pretreated data, a comprehensive feature vector and an importance score matrix of each feature are constructed; S4: the physical law of the power system is combined with artificial intelligence to construct a hybrid analysis model, and based on the comprehensive feature vector and the importance score matrix, hybrid analysis is performed to obtain an analysis result; and S5: the analysis result is converted into intuitive visual display and operable decision suggestions. The application realizes high-precision real-time analysis of distribution network line loss, and effectively improves the intelligentization and accuracy of line loss management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of line loss analysis, and particularly relates to a real-time line loss analysis method and system based on cross-modal measurement data fusion. BACKGROUND

[0002] Line loss of distribution network is a core index affecting the economy and safety of power system. With the large-scale application of smart meters, sensing terminals and various monitoring devices, massive, heterogeneous and real-time measurement data are generated in each link of power system, including multi-modal information such as current, voltage, power, temperature and fault alarm. This provides a data basis for more refined and intelligent line loss analysis and management, but also brings the following challenges:

[0003] Firstly, multi-source data have problems such as non-uniform format, different spatial and temporal resolution, many noise and abnormal points, and uneven quality. If these heterogeneous data cannot be effectively screened and fused, it is easy to lead to distorted line loss analysis results and even mislead decision-making. Secondly, traditional single algorithm often cannot simultaneously consider the advantages of power physical law and data-driven, and cannot effectively capture complex operating conditions and hidden line loss causes. In addition, how to intuitively display the analysis results and efficiently guide the actual operation, dispatching and management also needs to be innovatively solved. SUMMARY

[0004] In order to solve the above problems, the purpose of the present application is to provide a real-time line loss analysis method and system based on cross-modal measurement data fusion, which realizes high-precision real-time analysis of line loss of distribution network and effectively improves the intelligentization and accuracy of line loss management.

[0005] To achieve the above purpose, the present application adopts the following technical solutions:

[0006] A real-time line loss analysis method based on cross-modal measurement data fusion, comprising the following steps:

[0007] S1: Collecting multi-source heterogeneous data, and using an improved isolation forest algorithm to preliminarily screen the quality of the collected original data, obtaining an original data set and a corresponding data quality preliminary evaluation label;

[0008] S2: Preprocessing the original data set according to the data quality preliminary evaluation label to obtain preprocessed data;

[0009] S3: Constructing a comprehensive feature vector and an importance score matrix of each feature according to the preprocessed data;

[0010] S4: Deeply combining the physical law of power system with artificial intelligence to construct a hybrid analysis model, and based on the comprehensive feature vector and the importance score matrix, performing hybrid analysis to obtain an analysis result;

[0011] S5: transform the analysis results into intuitive visualizations and actionable decision recommendations.

[0012] Further, the multi-source heterogeneous data includes electrical operation data D scada of a SCADA system, high-frequency phasor data D pmu of a PMU system, user-side data D ami of a power utilization information collection system, infrared and visible light image data D img , environmental monitoring data D env , and static topology data D gis of a GIS system.

[0013] The electrical operation data includes voltage U scada , current I scada , active power P scada , and reactive power Q scada of each node of the power grid; the high-frequency phasor data of the PMU system includes voltage phasor V pmu , current phasor I pmu , phase angle θ pmu , and frequency f pmu of a substation and a line; the environmental monitoring data includes temperature T env , humidity H env , and wind speed W env .

[0014] The static topology data of the GIS system, i.e., power network topology Topo gis , contains static information of nodes, lines, and device parameters.

[0015] The above data is automatically synchronized and archived through an ETL interface, and the original data set D raw has the following structure:

[0016] D raw = {D scada , D pmu , D ami , D img , D env , D gis}.

[0017] Further, an improved isolation forest algorithm is used to preliminarily screen the collected original data, and the specific process is as follows:

[0018] According to the obtained original data set D , n is the number of samples, and each x i is a multi-modal feature splicing at the data collection time i, including D scada , D pmu , D ami , Dimg ,D env ,D gis ; each feature j has a corresponding data quality weight w j ;

[0019] Weighting features:

[0020] ;

[0021] where, denotes element-wise multiplication; denotes feature j data at collection time i;

[0022] In the improved isolation forest algorithm, when selecting a cutting point, if j is a time series feature, a time correlation constraint is introduced: let t be the time series corresponding to the current feature j, and the generation of the cutting point p considers not only the range of feature values but also the time neighborhood relationship:

[0023] ;

[0024] where r(t) is the autocorrelation function of time; a is an adjustment coefficient; , and are the minimum and maximum values of feature j in the sample set, respectively; U represents a uniform distribution;

[0025] In the improved isolation forest algorithm, the generation of each tree in the tree construction process is as follows:

[0026] Construction input: sample subset , is the jth data in the sample subset, n is the data quantity; the current tree depth l , maximum depth l max Output: an isolation tree;

[0027] Termination condition: if |X|≤1 or l ≥ l max, return a leaf node, and record the sample number |X| in the node;

[0028] Select a feature:

[0029] Randomly select feature j with probability , where Avgweight j is the average weight of feature j on X;

[0030] Select a cutting point:

[0031] If j is a time series feature, select p according to the time-aware cutting point generation method described above; otherwise, uniformly randomly select p between and ;

[0032] Splitting data to get left sub-tree data X left , right sub-tree data X right :

[0033] ;

[0034] Recursive construction of sub-trees, including left sub-tree and right sub-tree:

[0035] Weighted path length is adopted:

[0036] At node s, if splitting according to feature j, the value of accumulated path length when passing through node s is , where w j (x') is the weight of current sample x' on feature j;

[0037] Weighted path length calculation:

[0038] ;

[0039] where Path(x') is the path of sample x' in single weighted isolation tree, i.e. the set of all splitting steps experienced from root node to its leaf node; h dw (x') is the weighted path length; is the weight value of the path from j to s under sample x';

[0040] After generating the forest of t' trees, the weighted average path length of sample x' is calculated :

[0041] ;

[0042] where, is the weighted path length of sample x' on the kth weighted isolation tree;

[0043] Normalization of anomaly score:

[0044] ;

[0045] where c(n'') is the normalization factor of node n'';

[0046] Get the anomaly score s dw of each sample, and then set threshold τ to determine anomaly:

[0047]

[0048] where Q is the final anomaly discrimination label.

[0049] Further, preprocessing, specifically as follows:

[0050] According to the anomaly score and the anomaly discrimination label, all data samples marked as severe anomalies are directly removed, and the processed data is D filtered :

[0051] ;

[0052] wherein, is the tolerable floating space for retaining part of the boundary samples to enhance the model robustness; Q i is the anomaly discrimination label of the data x i ;

[0053] Detect missing values for each data x filtered and each feature x i in D ij , and use mean imputation;

[0054] Only when the anomaly score s dw is in the adjustment interval, replace the anomaly value with the context-based mean; and normalize to obtain the pre-processed data set D pre .

[0055] Further, according to the pre-processed data, a comprehensive feature vector and an importance score matrix of each feature are constructed, as follows:

[0056] According to the processed data set D pre , for time series numerical data, statistical features including mean, extreme value, variance, skewness, and kurtosis are extracted; for image data, convolutional neural network model is used to extract key visual features including fault hotspots and structural change features; for environmental and geographic information, trend features are extracted through a statistical window; and for static or categorical data, vectorization processing is performed, and all normalized data are fused into a single comprehensive feature vector in a splicing manner;

[0057] After obtaining the comprehensive feature vector, the specific contribution of each type of feature to the subsequent power system line loss prediction is quantified.

[0058] Further, the physical law of the power system is deeply combined with artificial intelligence to construct a hybrid analysis model, as follows:

[0059] Define a physical model module including a first physical model and a second physical model, as follows:

[0060] First physical model: for the i'th sample, calculate the theoretical line loss based on physical quantity calculation theory :

[0061]

[0062] wherein, is the k'th phase current of the i'th sample, is the corresponding resistance; K is the number of phases;

[0063] Second physical model: define the voltage out-of-limit, current out-of-limit state quantity function, used to constrain the subsequent AI branch output:

[0064]

[0065] wherein, is the measured voltage value of the i'th sample; is the normal upper and lower bound of the voltage under physical constraints; is the indicator function; is the voltage state constraint function, which returns 0 normally and 1 when out-of-limit, used for loss function or early warning output;

[0066] The comprehensive feature vector z is weighted according to the importance vector m to obtain the enhanced feature :

[0067] ;

[0068] wherein, z i′ is the comprehensive feature vector of the i'th sample; m is the feature importance weight vector; and is the element-wise multiplication;

[0069] Then the physical quantity is added as a feature to form a hybrid feature vector :

[0070] ;

[0071] wherein, represents vector splicing; p i′ is the physical feature vector of the i'th sample;

[0072] A hybrid neural network model is constructed, including a dual-branch neural network of a physical branch and an AI branch:

[0073] The physical branch inputs the physical feature p i′ , generates a physical latent representation phys through a fully connected layer FC :

[0074]

[0075] The AI branch inputs the importance-weighted feature , generates an AI latent representation through a multi-layer perceptron MLP AI :

[0076]

[0077] Fusion layer: concatenate the outputs of two branches, generate the final task-related representation through a joint fully connected layer

[0078]

[0079] where W f , b f are the weight and bias of the fully connected layer respectively; ReLU is the activation function

[0080] The output layer gets the final line loss prediction value , anomaly detection probability and voltage prediction value

[0081]

[0082]

[0083]

[0084] where W loss , b loss are the weight vector and bias of the line loss prediction output layer respectively; W o , b o are the weight and bias of the anomaly detection output layer respectively; W volt , b volt are the weight and bias of the voltage prediction output layer respectively.

[0085] Further, the loss function of the hybrid analysis model includes task loss, physical consistency loss and physical constraint loss, as follows:

[0086] Task loss L task

[0087]

[0088] where y i′ is the true value of the i′ sample; n′ is the number of samples

[0089] Physical consistency loss L consist

[0090]

[0091] Physical constraint loss L constraint

[0092] ⋅

[0093] where λ​​​​​​​​​​​voltage To physically constrain the weights;

[0094] Final loss function:

[0095] L = L task + λ1L consist + λ2L constraint ;

[0096] Where λ1, λ2 are adjustable hyperparameters, controlling the strength of physical law embedding.

[0097] Further, the analysis results are converted into intuitive visualizations and actionable decision recommendations, as follows:

[0098] For mixed analysis model analysis results, including line loss prediction, anomaly detection, voltage overrun, structured induction and template-driven automatic expression are performed;

[0099] Taking line loss prediction, anomaly probability, and voltage overrun indicators as input, a multi-level heat map is dynamically generated and superimposed on the power grid geographic information map to visually see the risk distribution of different areas and lines;

[0100] Deep reinforcement learning is introduced to automatically generate the best response strategy based on the analysis results of the mixed analysis model, historical data, and power network simulation.

[0101] Further, deep reinforcement learning is introduced to automatically generate the best response strategy based on the analysis results of the mixed analysis model, historical data, and power network simulation, as follows:

[0102] Based on historical data and power grid topology, a simulation environment is constructed, which can simulate the state of the power grid;

[0103] State transition: the state of the power grid s t is transferred to s t under action a t+1 The dynamic process is realized by the simulator:

[0104] s t+1 = E(s t , a t );

[0105] Where E is the simulation engine;

[0106] The state vector is represented as:

[0107] ;

[0108] Where, are the line loss prediction value, anomaly detection probability, and voltage prediction value at the current time t, respectively; I t is the line current, and P tP t V t λ loss,t H anom,t P P

[0109] a t The multi-level optimization response strategy includes:

[0110] Immediate action Load adjustment, switch operation, and capacitor switching;

[0111] Short-term maintenance Issuing inspection work orders and temporary inspection equipment list L chec ;

[0112] Long-term transformation Applying for technical transformation budget and planning to upgrade region Z upgrade ;

[0113] The action space is a hierarchical structure:

[0114] ;

[0115] The reward function r t is defined as:

[0116] ;

[0117] where α1, α2, and α3 are weight coefficients; is an out-of-limit indicator function; Cost(a t ) is the action cost; V ref is the voltage reference value; ΔV is the voltage allowable deviation range; is the predicted line loss value at time t; is the abnormal event discrimination output at time t; is the predicted voltage value at time t; β' is the weight coefficient of the control action cost term;

[0118] A hierarchical reinforcement learning framework combined with the Actor-Critic algorithm is adopted:

[0119] Critic network: evaluate state value V(s t ; θv) or action value Q(s t , a t ; θ q ); where θv and θq are the parameters of the state value function V and the action value function Q, respectively;

[0120] Actor network: output hierarchical action strategy π(a t | s t; theta π );

[0121] Optimization goal: maximize cumulative discounted reward :

[0122]

[0123] where gamma is a discount factor; is an expected function; T is the total time;

[0124] Update formula:

[0125] ;

[0126] where, is the gradient of the objective function J with respect to the policy parameter theta; is the log policy gradient, representing the gradient of the log probability of selecting action a t in state s t with respect to the parameter theta;

[0127] A(s t ,a t ) is the advantage function, estimated by TD-error:

[0128] ;

[0129] where V(s t ) is the value function estimate of the current state s t .

[0130] A real-time line loss analysis system based on cross-modal measurement data fusion, comprising a processor, a memory and a computer program stored in the memory, when the processor executes the computer program, specifically executes the steps of a real-time line loss analysis method based on cross-modal measurement data fusion.

[0131] The present application has the following beneficial effects:

[0132] 1、The present application innovatively introduces modal distinction, feature weighting and data quality weight through the multi-source heterogeneous data structure of the power system, standardizes different sampling frequencies and heterogeneous feature data in layers, and through weight distribution and weighted path length calculation, can effectively suppress the interference of low-quality data in actual operation, and significantly reduce false positives and false negatives.

[0133] 2、The application explicitly embeds the physical law of the power system in the AI modeling process, such as line loss theory calculation and state constraint criterion, which not only ensures the credibility and safety of the model output in the physical layer, avoids abnormal results without physical meaning, but also realizes the high integration of data-driven and knowledge-driven, and fully utilizes the domain prior knowledge; and through feature importance weighting and mixed physical feature splicing, the model can adaptively highlight the information most relevant to business value. The weighted comprehensive feature vector improves the perception ability of the model to abnormal and key business features, effectively improves the accuracy and robustness of analysis;

[0134] 3、The application not only presents in an intuitive visual form, including multi-dimensional interactive heat map and core business index large screen, but also generates an operable suggestion list through automatic text generation, providing all-round, hierarchical and personalized decision support for operation and management personnel, and responding to potential risks in real time, proposing immediate, short-term and long-term optimization suggestions, greatly improving the scientificity and intelligent level of distribution network line loss management and operation decision. BRIEF DESCRIPTION OF DRAWINGS

[0135] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0136] The application will be further described in detail below in combination with the drawings and specific embodiments:

[0137] REFERENCE Figure 1 In this embodiment, a real-time line loss analysis method based on cross-modal measurement data fusion is provided, including the following steps:

[0138] S1: Collect multi-source heterogeneous data, and use an improved isolation forest algorithm to preliminarily screen the collected raw data, obtain a raw data set, and a corresponding data quality preliminary evaluation label;

[0139] S2: According to the data quality preliminary evaluation label, the raw data set is preprocessed to obtain the preprocessed data;

[0140] S3: According to the preprocessed data, a comprehensive feature vector and an importance score matrix of each feature are constructed;

[0141] S4: The physical law of the power system is combined with artificial intelligence to construct a hybrid analysis model, and based on the comprehensive feature vector and the importance score matrix, hybrid analysis is performed to obtain an analysis result;

[0142] S5: The analysis result is converted into intuitive visual display and operable decision suggestion.

[0143] In this embodiment, the multi-source heterogeneous data includes the electrical operation data Dscada (voltage, current, power, etc., sampling period 5-15 seconds), high-frequency phasor data D of the PMU system pmu (millisecond-level sampling), user-side data D of the power consumption information collection system ami (15-minute sampling), infrared and visible light image data D img (hourly or daily sampling), environmental monitoring data D env (minute-level sampling) and static topology data D of the GIS system gis ;

[0144] The electrical operation data include voltages U scada , currents I scada , active powers P scada , and reactive powers Q scada (sampling period 5-15 seconds); the high-frequency phasor data of the PMU system include voltage phasors V pmu , current phasors I pmu , phase angles θ pmu , and frequencies f pmu ; the environmental monitoring data include temperatures T env , humidities H env , and wind speeds W env ;

[0145] The static topology data of the GIS system, i.e., the power network topology Topo gis , contains static information of nodes, lines, and device parameters;

[0146] The above data are automatically synchronized and archived through an ETL interface, and the original data set D raw has the following structure:

[0147] D raw ={D scada ,D pmu ,D ami ,D img ,D env ,D gis}。

[0148] In this embodiment, the improved isolation forest algorithm is used to preliminarily screen the collected original data, and the specific process is as follows:

[0149] According to the obtained original data set , n is the number of samples, and each x i is a multi-modal feature splice at the data collection time i, including D scada ,D pmu ,D ami ,D img ,Denv D gis ; each feature j has a corresponding data quality weight w j ;

[0150] Weighting features:

[0151] ;

[0152] wherein, denotes element-wise multiplication; denotes feature j data at collection time i;

[0153] In the improved isolation forest algorithm, when selecting a cutting point, if j is a time series feature, a time correlation constraint is introduced: assuming that the current feature j corresponds to a time series t, the generation of the cutting point p considers not only the feature value range but also the time neighborhood relationship:

[0154] ;

[0155] wherein, r(t) is a time autocorrelation function (such as ), t' is the current point time, β is a time correlation decay coefficient); α is an adjustment coefficient (such as α = 0.1); 、 are the minimum and maximum values of feature j in the sample set respectively; U represents a uniform distribution;

[0156] In the improved isolation forest algorithm, the generation of each tree in the tree construction process is as follows:

[0157] Construction input: sample subset , is the jth data in the sample subset, n is the data quantity; the current tree depth l , the maximum depth l max Output: an isolation tree;

[0158] Termination condition: if ∣X∣≤1 or l ≥ l max, return a leaf node, and record the sample number ∣X∣ in the node;

[0159] Selecting a feature:

[0160] Randomly select a feature j with a probability , wherein Avgweight j is the average weight of the feature on X;

[0161] Selecting a cutting point:

[0162] If j is a time series feature (such as voltage and current), the p is selected according to the above time-aware cutting point generation method; otherwise, in and p is uniformly randomly selected among

[0163] Split the data to get left child data X left and right child data X right :

[0164] ;

[0165] Recursively build the child trees, including the left child and the right child:

[0166] In traditional isolation forest, the path length h(x') is defined as the number of edges from the root node to the leaf node where the sample x' is located.

[0167] In this scheme, the weighted path length is adopted:

[0168] At node s, if the data is split according to feature j, the value of the accumulated path length when passing through this node is where w j (x') is the weight of the current sample x' on feature j;

[0169] The calculation of the weighted path length:

[0170] ;

[0171] where Path(x') is the path of sample x' in a single weighted isolation tree, i.e., the set of all split steps from the root node to its leaf node; h dw (x') is the weighted path length; is the weight value of the path from j to s under sample x';

[0172] After generating a forest of t' trees, the weighted average path length of sample x' is calculated as :

[0173] ;

[0174] where, is the weighted path length of sample x' on the kth weighted isolation tree;

[0175] The normalized anomaly score:

[0176] ;

[0177] where c(n'') is the normalization factor of node n'';

[0178] Get the anomaly score s dw of each sample, and then set the threshold τ (τ = 0.6) to determine the anomaly:

[0179]

[0180] wherein Q is the final anomaly discrimination label.

[0181] In this embodiment, the preprocessing is specifically as follows:

[0182] According to the anomaly score and the anomaly discrimination label, all data samples labeled as severe anomaly are directly removed to avoid the influence of extreme values on subsequent modeling, and the processed data is D filtered :

[0183] ;

[0184] wherein, is the tolerable floating space for retaining part of the boundary samples to enhance the robustness of the model; Q i is the anomaly discrimination label of the data x i ;

[0185] For each data x filtered and each feature x i in D ij , missing values (such as NaN, null, invalid encoding) are detected, and mean imputation is adopted;

[0186] Only when the anomaly score s dw is located in the adjustment interval, the anomaly value is replaced with the context-based mean value; and normalization processing is performed to obtain the preprocessed data set D pre .

[0187] In this embodiment, according to the preprocessed data, a comprehensive feature vector and an importance score matrix of each feature are constructed, specifically as follows:

[0188] According to the processed data set D pre , for SCADA, PMU and other time series numerical data, statistical features including mean, extreme value, variance, skewness and kurtosis are extracted; for infrared, visible light and other image data, convolutional neural network model is adopted to extract key visual features including fault hotspots and structural change features; for environmental and geographical information, trend features are extracted through a statistical window; static or categorical data are subjected to vectorization processing (such as One-Hot), and all the normalized data are fused into a single comprehensive feature vector in a splicing manner;

[0189] After obtaining the comprehensive feature vector, the specific contribution of each type of feature to the subsequent line loss prediction of the power system is quantified, and a tree model (such as random forest, XGBoost, etc.) is adopted to automatically output a feature split gain index to form a feature importance ranking.

[0190] In this embodiment, the physical law of power system is deeply combined with artificial intelligence to construct a hybrid analysis model, which is specifically as follows:

[0191] A physical model module is defined, including a first physical model and a second physical model, which are specifically as follows:

[0192] The first physical model (line loss physical calculation): for the i'th sample (such as a line / area), the theoretical line loss is calculated based on physical quantities (current, resistance, voltage, etc.) :

[0193]

[0194] wherein, is the k'th phase current (or total current) of the sample, is the corresponding resistance; K is the number of phases;

[0195] The second physical model (state constraint function): define the voltage out-of-limit and current out-of-limit state quantity function, which is used to constrain the subsequent AI branch output:

[0196]

[0197] wherein, is the voltage value measured by the i'th sample; is the normal upper and lower bound of voltage under physical constraint; is an indicator function; is the voltage state constraint function, which returns 0 normally and 1 when out-of-limit, which can be used in loss function or early warning output;

[0198] The comprehensive feature vector z is weighted according to the importance vector m to obtain the enhanced feature :

[0199] ;

[0200] wherein, z i′ is the comprehensive feature vector of the i'th sample (from S3, the feature after multi-modal splicing, the dimension is d); m is the feature importance weight vector (dimension d, reflecting the importance of each dimension feature in the subsequent task); is the element-wise multiplication;

[0201] Then the physical quantity is added as a feature to form a hybrid feature vector :

[0202] ;

[0203] wherein, represents vector splicing; p i′ is the physical feature vector of the i'th sample (such as , Vi′ , etc.);

[0204] The mixed neural network model is a double-branch neural network including a physical branch and an AI branch:

[0205] The physical branch inputs physical features p i′ , and generates a physical latent representation through a fully connected layer FC phys :

[0206]

[0207] The AI branch inputs importance-weighted features , and generates an AI latent representation through a multi-layer perceptron MLP AI :

[0208]

[0209] The fusion layer splices the outputs of the two branches, and generates a final task-related representation through a joint fully connected layer:

[0210] ;

[0211] wherein W f , b f are the weight and bias of the fully connected layer respectively; and ReLU is an activation function;

[0212] The output layer obtains a final line loss prediction value , an anomaly detection probability , and a voltage prediction value :

[0213] ;

[0214] ;

[0215] ;

[0216] wherein W loss , b loss are the weight vector and bias of the line loss prediction output layer respectively; W o , b o are the weight and bias of the anomaly detection output layer respectively; and W volt , b volt are the weight and bias of the voltage prediction output layer respectively.

[0217] In this embodiment, the loss function of the mixed analysis model includes a task loss, a physical consistency loss, and a physical constraint loss, and is specifically as follows:​​​

[0218] Task loss L task :

[0219] ;

[0220] wherein y i′ is the true value of the i' sample; n' is the number of samples;

[0221] Physical consistency loss L consist :

[0222] ;

[0223] Physical constraint loss L constraint :

[0224] .

[0225] wherein λ voltage is the physical constraint weight;

[0226] Final loss function L:

[0227] L = L task + λ1L consist + λ2L constraint ;

[0228] wherein λ1, λ2 are adjustable hyperparameters, controlling the embedding strength of physical laws.

[0229] In the embodiment, the analysis results are converted into intuitive visual display and operable decision suggestions, as follows:

[0230] For the analysis results of the mixed analysis model, including line loss prediction, anomaly detection, and voltage overrun, structured induction and template-driven automatic expression are performed, and complex time series prediction, abnormal distribution, and attribution information are converted into textual analysis reports that are easy for business personnel to understand;

[0231] Taking line loss prediction, anomaly probability, and voltage overrun indicators as inputs, a multi-level heat map is dynamically generated and superimposed on the power grid geographic information map to visually see the risk distribution of different areas and lines;

[0232] Deep reinforcement learning is introduced, and based on the analysis results of the mixed analysis model, historical data, and power network simulation, the best response strategy is automatically generated.

[0233] In the embodiment, deep reinforcement learning is introduced, and based on the analysis results of the mixed analysis model, historical data, and power network simulation, the best response strategy is automatically generated, as follows:

[0234] Based on historical data and power grid topology, a simulation environment (e.g., power system simulator based on OpenDSS or MATLAB / Simulink) is constructed to simulate the power grid operating state.

[0235] State transition: Power grid state s t In action a t transitions to s t+1 The dynamic process is implemented by the simulator:

[0236] s t+1 =E(s t ,a t );

[0237] Where E is the simulation engine;

[0238] State s t integrates the following information:

[0239] Model analysis results: line loss prediction value, anomaly detection probability, voltage prediction value at the current time;

[0240] Physical state: line current It, active power Pt, node voltage Vt, load rate λt;

[0241] Historical state: historical data within the sliding window (e.g., state sequence of the past few time points);

[0242] Heat map aggregation indicators: line loss heat value H loss,t , anomaly heat value H anom,t , etc. of the current regions.

[0243] The state vector is represented as:

[0244]

[0245] Where, are the line loss prediction value, anomaly detection probability, and voltage prediction value at the current time t; I t is the line current, P t is the active power, V t is the node voltage, and λ t is the load rate; H loss,t , H anom,t represent the line loss heat value and anomaly heat value of the current regions, respectively.

[0246] Action a t is defined as a multi-level optimization response strategy, including:

[0247] Instant action : Load adjustment (adjustment ratio δ), switch operation (open / close), capacitor switching;

[0248] ​Short-term maintenance : dispatch inspection work order (object list Linsp), inspection equipment list L chec ;

[0249] Long-term transformation : apply for technical transformation budget, plan upgrade area Z upgrade ;

[0250] The action space is a hierarchical structure:

[0251] ;

[0252] Define the reward function r t :

[0253] ;

[0254] Wherein, α1, α2, α3 are weight coefficients; is an out-of-limit indication function; Cost(a t ) is the action cost; V ref is the voltage reference value; ΔV is the voltage allowable deviation range; is the predicted line loss value at time t; is the abnormal event discrimination output at time t; is the predicted voltage value at time t; β' is the weight coefficient of the control action cost term;

[0255] A hierarchical reinforcement learning framework (such as Option-Critic or MAXQ) is adopted in combination with the Actor-Critic algorithm:

[0256] Critic network: evaluate state value V(s t ;θv) or action value Q(s t ,a t ;θ q ); wherein, θv, θq are parameters of state value function V and action value function Q respectively;

[0257] Actor network: output hierarchical action policy π(a t ∣s t ;θ π );

[0258] Optimization goal: maximize cumulative discounted reward :

[0259]

[0260] Wherein, γ is the discount factor; is the expected function; T is the total time;

[0261] Update formula:

[0262] ;

[0263] wherein, is the gradient of the objective function J with respect to the policy parameters θ; is the log policy gradient, representing the gradient of the log probability of selecting action a t in state s t with respect to the parameters θ;

[0264] A(s t ,a t ) is the advantage function, estimated by the TD-error:

[0265] ;

[0266] wherein V(s t ) is the value function estimate for the current state s t .

[0267] A real-time line loss analysis system based on cross-modal measurement data fusion, comprising a processor, a memory and a computer program stored in the memory, when the processor executes the computer program, specifically executes the steps of a real-time line loss analysis method based on cross-modal measurement data fusion

[0268] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROMs, optical storage media, etc.) having computer usable program code embodied therein.

[0269] The present application is described in reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowchart and / or block diagram of the flow or multiple flows and / or blocks. Figure 1 The flowchart and / or block diagram of the flowchart and / or block diagram of the flow or multiple flows and / or blocks. Figure 1 The flowchart and / or block diagram of the flowchart and / or block diagram of the flow or multiple flows and / or blocks.

[0270] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 The functions described can be implemented in one or more flow(s), and / or block(s) and combinations of flow(s) and / or block(s) Figure 1 The functions described can be implemented in one or more flow(s), and / or block(s) and combinations of flow(s) and / or block(s)

[0271] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow Figure 1 The functions described can be implemented in one or more flow(s), and / or block(s) and combinations of flow(s) and / or block(s) Figure 1 The functions described can be implemented in one or more flow(s), and / or block(s) and combinations of flow(s) and / or block(s)

[0272] The above descriptions are only preferred embodiments of the present application and are not intended to limit the present application to other forms. Any person skilled in the art can make modifications or alterations to the above-described embodiments without departing from the technical scope of the present application. Any simple modifications, equivalent changes and alterations made to the above-described embodiments based on the technical essence of the present application without departing from the technical scope of the present application are still within the protection scope of the present application.

Claims

1. A real-time line loss analysis method based on cross-modal measurement data fusion, characterized in that, The method comprises the following steps: S1: Collecting multi-source heterogeneous data, and using an improved isolation forest algorithm to perform preliminary quality screening on the collected raw data, obtaining a raw data set and a corresponding preliminary data quality evaluation label; S2: According to the data quality preliminary evaluation label, the raw data set is preprocessed to obtain preprocessed data; S3: According to the preprocessed data, a comprehensive feature vector and an importance score matrix of each feature are constructed; S4: The physical law of the power system is deeply combined with artificial intelligence to construct a hybrid analysis model, and based on the comprehensive feature vector and the importance score matrix, hybrid analysis is performed to obtain an analysis result; S5: The analysis result is converted into intuitive visual display and operable decision suggestions; The hybrid analysis model is constructed by deeply combining the physical law of the power system with artificial intelligence, and specifically as follows: A physical model module is defined, including a first physical model and a second physical model, and specifically: First physical model: For the i'th sample, theoretical line losses are calculated based on physical quantities : ； wherein, is the k'th phase current for the i'th sample, is the corresponding resistance; K is the number of phases; The second physical model defines a voltage out-of-limit and current out-of-limit state quantity function for constraining subsequent AI branch output: ； wherein, Vmeas(i) is the voltage value measured for the i-th sample; Vmin(i) and Vmax(i) are the normal lower and upper bounds for the voltage under physical constraints; I(i) is an indicator function; Vcon(i) is a voltage state constraint function, returning 0 for normal and 1 for out-of-limits, used in the loss function or early warning output; The enhanced features are obtained by weighting the comprehensive feature vector z by the importance vector m : ； wherein z i′ is the integrated feature vector of the i'th sample; m is the feature importance weight vector; and is the element-wise multiplication. Then, the physical quantity is added as a feature to form a hybrid feature vector : ； wherein, represents vector concatenation; p i′ is the physical feature vector of the i'th sample; A hybrid neural network model is constructed, including a double-branch neural network of a physical branch and an AI branch: the physical branch, input physical features p i′ , through a fully connected layer FC phys generating a physical latent representation : ； The AI branch inputs importance-weighted features , by a multi-layer perceptron, MLP AI generates an AI latent representation : ； fusion layer: concatenating the outputs of the two branches through a joint fully connected layer to generate the final task-relevant representation : ； where W f , b f are the weights and biases of the fully connected layer, respectively; ReLU is the activation function. The output layer obtains the final line loss prediction value , anomaly detection probability and voltage prediction value : ； ； ； wherein W loss , b loss are the weight vector and bias of the line loss prediction output layer, respectively; W o , b o are the weight and bias of the anomaly detection output layer, respectively; W volt , b volt are the weight and bias of the voltage prediction output layer, respectively.

2. The real-time line loss analysis method based on cross-modality measurement data fusion according to claim 1, characterized in that, The multi-source heterogeneous data includes electrical operation data D of a SCADA system scada , high-frequency phasor data D of a PMU system pmu , user-side data D of a power utilization information collection system ami , infrared and visible light image data D img , environmental monitoring data D env , and static topology data D of a GIS system gis ; The electrical operation data includes voltage U scada , current I scada , active power P scada , and reactive power Q scada of each node of the power grid; the high-frequency phasor data of the PMU system includes voltage phasor V pmu , current phasor I pmu , phase angle θ pmu , and frequency f pmu of the transformer substation and the line; and the environmental monitoring data includes temperature T env , humidity H env , and wind speed W env . Static topological data of GIS system, i.e. power network topology Topo gis , contains node, line, device parameter static information; The above data is automatically synchronized and archived through the ETL interface, and the original data set D raw The structure is as follows: D raw ={D scada ,D pmu ,D ami ,D img ,D env ,D gis}。 3. The real-time line loss analysis method based on cross-modality measurement data fusion according to claim 2, characterized in that, The improved isolation forest algorithm is used to perform preliminary quality screening on the collected raw data, and specifically as follows: According to the obtained original data set , n is the number of samples, each x i is a multi-modal feature splicing at data collection moment i, including D scada , D pmu , D ami , D img , D env , D gis ; each feature j has a corresponding data quality weight w j ; The features are weighted: ； wherein, represents element-wise multiplication; represents feature j data at collection time i; In the selection of the cutting point, if j is a time sequence feature, a time correlation constraint is introduced: assuming that the current feature j corresponds to a time sequence t, the generation of the cutting point p considers not only the feature value range but also the time neighborhood relationship: ； Wherein, r(t) is a time autocorrelation function; a is an adjustment coefficient; are the minimum and maximum values of feature j in the sample set, respectively; U denotes a uniform distribution; In the construction process of the tree, each tree is generated as follows: Construct input: sample subset , For the jth data in the sample subset, n is the data amount; the current tree depth l, the maximum depth l max ; Output: an isolated tree; Termination condition: if |X|≤1 or l≥lmax, return the leaf node, and record the sample number |X| in the node; Selecting a feature: with probability randomly select a feature j, where Avgweight j is the average weight of feature j over X; Selecting a cutting point: If j is a temporal feature, then p is selected according to the time-aware cutpoint generation method described above; otherwise, p is uniformly randomly selected between and ​ Splitting the data to get left sub-tree data X left , right sub-tree data X right : ； Recursively constructing a subtree, including a left subtree and a right subtree: Using a weighted path length: At node s, the value of the accumulated path length through node s is where w j (x') is the weight of the current sample x' on feature j. Weighted path length calculation: ； wherein Path(x') is the path of the sample x' in the single weighted isolation tree, i.e. the set of all split steps experienced from the root node to its leaf node; h dw (x') is the weighted path length; is the weight value of the path from j to s under the sample x'. After generating the forest of t' trees, the weighted average path length is computed for the sample x' : ； wherein, is the weighted path length of sample x' on the kth weighted isolation tree; Normalization of abnormal score: ； Wherein, c(n'') is the normalization factor of node n''; obtain an anomaly score s for each sample dw and then set a threshold τ to decide anomalies: ； Wherein, Q is the final abnormality judgment label.

4. The real-time line loss analysis method based on cross-modality measurement data fusion according to claim 3, characterized in that, The preprocessing is specifically as follows: According to the anomaly score and the anomaly discrimination label, all data samples labeled as severe anomaly are directly removed, and the processed data is D filtered : D filtered = {x i | Q i = 0 or Q i = 1, s dw ≤ τ+ } wherein, is a tolerable up-floating space for reserving partial boundary samples to enhance model robustness; Q i is an abnormality discrimination label of data x i ; For D filtered each data x i and each feature x ij missingness is detected and mean imputation is employed; only if the anomaly score s dw When in the adjustment interval, replace the anomaly value with the context-based mean; and normalize to obtain the preprocessed data set D pre .

5. The real-time line loss analysis method based on cross-modality measurement data fusion according to claim 4, characterized in that, According to the preprocessed data, a comprehensive feature vector and an importance score matrix of each feature are constructed, and specifically as follows: According to the processed data set D pre For time series numerical data, the extracted statistical features include mean, extreme value, variance, skewness, kurtosis; for image data, the convolutional neural network model is used to extract key visual features, including fault hot spots and structural change features; for environmental and geographic information, trend features are extracted through a statistical window; The static or category data is vectorized, and all the normalized data is fused into a single comprehensive feature vector in a splicing manner; After obtaining the comprehensive feature vector, the specific contribution of each type of feature to the subsequent line loss prediction of the power system is quantified.

6. The real-time line loss analysis method based on cross-modality measurement data fusion according to claim 1, characterized in that, The loss function of the hybrid analysis model includes task loss, physical consistency loss and physical constraint loss, and specifically as follows: Task loss L task : ； where y i′ is the true value of the i'th sample; n' is the number of samples; Physical consistency loss L consist : ； Physical constraint loss L constraint : ； where λ voltage is a physical constraint weight; Final loss function: L = L task + λ1L consist + λ2L constraint ; Wherein, λ1 and λ2 are adjustable hyperparameters, controlling the embedding strength of the physical law.

7. The real-time line loss analysis method based on cross-modality measurement data fusion according to claim 1, characterized in that, The analysis result is converted into intuitive visual display and operable decision suggestions, and specifically as follows: The hybrid analysis model analysis result, including line loss prediction, abnormality detection and voltage out-of-limit, is structured and automatically expressed in a template-driven manner; Taking line loss prediction, abnormality probability and voltage out-of-limit indicators as inputs, a multi-level heat map is dynamically generated and superimposed on the power grid geographic information map to visually see the risk distribution of different areas and lines. The deep reinforcement learning is introduced to automatically generate the optimal response strategy based on the analysis result of the hybrid analysis model, historical data and power network simulation.

8. The real-time line loss analysis method based on cross-modality measurement data fusion according to claim 7, characterized in that, The deep reinforcement learning is introduced to automatically generate the optimal response strategy based on the analysis result of the hybrid analysis model, historical data and power network simulation, and the specific implementation is as follows: Based on historical data and power grid topology, a simulation environment is constructed to simulate the operation state of the power grid. State transition: grid state s t In action a t Down to s t+1 The dynamic process is implemented by the simulator: s t+1 =E(s t ,a t ); The simulation engine is denoted as E. The state vector is denoted as: ； wherein, are the line loss prediction value, the abnormality detection probability, and the voltage prediction value at the current time t, respectively; I t is the line current, P t is the active power, V t is the node voltage, λ t is the load rate; H loss,t , H anom,t respectively represent the line loss thermal value and the abnormality thermal value of the current region; Action a t Defined as a multi-level optimization response strategy, including: Instant action : load adjustment, switch operation, capacitor switching Short term maintenance : dispatching patrol inspection work order, temporary inspection equipment list L chec ; Long-term renovation : Apply for budget for technical renovation, plan to upgrade area Z upgrade ; The action space is a hierarchical structure. ； Defining the reward function r t : ； wherein, a1, a2, a3 are weight coefficients; is an out-of-limit indication function; Cost(a t ) is an action cost; V ref is a voltage reference value; and AV is a voltage allowable deviation range; is a predicted line loss value at time t; is an abnormal event discrimination output at time t; is a predicted voltage value at time t; and β' is a weight coefficient of a control action cost term; The hierarchical reinforcement learning framework is adopted in combination with the Actor-Critic algorithm. Critic network: evaluates the state value V(s t ; θv) or the action value Q(s t , a t ; θ q ); where θv, θq are parameters of the state value function V and the action value function Q, respectively; Actor network: output hierarchical action policy p(a t | s t ; Q π ) Optimization goal: maximize cumulative discounted reward : ； where γ is a discount factor; is the expected function; T is the total time; The update formula is as follows: ； wherein, is the gradient of the objective function J with respect to the policy parameters Q; is the log policy gradient, representing the gradient of the log probability of selecting action a t in state s t with respect to the parameters Q; A(s t ,a t ) is the advantage function, estimated by the TD-error: ； where V(s t ) is the value function estimate for the current state s t .

9. A real-time line loss analysis system based on cross-modality measurement data fusion, characterized in that, The computer program comprises a processor, a memory and a computer program stored in the memory, and the processor executes the computer program to specifically execute the steps of the real-time line loss analysis method based on cross-modal measurement data fusion according to any one of claims 1-8.

Images