An AI-based optimization method for identifying and addressing energy consumption anomalies in power distribution substations.

Abstract

This invention relates to the field of intelligent power distribution network anomaly monitoring, and particularly to an AI-based optimization method for identifying and addressing energy consumption anomalies in power distribution substations. The method includes: establishing a sliding window to collect power and temperature / humidity data to generate a multimodal feature sequence; mapping a physical topology graph and extracting spatially correlated topological feature vectors using a graph attention network; calculating Copula entropy to remove temperature / humidity dependence; using a sparse Gaussian process to fit decoupled features to construct a dynamic energy consumption baseline with predictive variance; comparing measured data with the baseline to mark anomalous nodes; transforming the topology graph into an absorbing Markov chain transition matrix to calculate the absorption probability and output the source tracing result; and using the result as a feedback sample to update the network model parameters. This invention can eliminate meteorological variation interference, improve the noise resistance and stability of the monitoring system, and achieve high-confidence identification and global probabilistic topology tracing of equipment anomalies.

Description

Technical Field

[0001] This invention relates to the field of anomaly monitoring in smart power distribution networks, and in particular to an optimization method for identifying anomalies in energy consumption of power distribution substations based on AI algorithms. Background Technology

[0002] The power consumption of electrical equipment in substations is not only driven by the end-user electrical load, but also has a complex nonlinear coupling with meteorological conditions such as temperature and humidity. Especially during extreme weather changes, equipment often experiences reasonable spikes in energy consumption. To prevent potential hazards such as insulation aging, the industry urgently needs to conduct precise monitoring and in-depth analysis of multi-dimensional energy consumption data to identify subtle distortions in equipment status in advance and achieve accurate predictive maintenance.

[0003] However, existing systems mostly rely on static threshold alarms, which cannot effectively decouple meteorological interference. This makes it difficult for the system to separate reasonable energy consumption spikes caused by extreme weather changes, and it is easy to misjudge them as real equipment failures, thus causing false alarms caused by sudden environmental changes. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an AI-based method for identifying and optimizing energy consumption anomalies in power distribution substations, aiming to improve the problem of false alarms caused by meteorological interference in existing static threshold monitoring.

[0005] This invention provides the following technical solution: a method for identifying and optimizing energy consumption anomalies in power distribution substations based on AI algorithms, comprising:

[0006] S1. Establish a sliding time window, collect the operating power data and ambient temperature and humidity data of each electrical device in the substation, and perform spatiotemporal alignment and interpolation preprocessing to generate multimodal feature sequences.

[0007] S2. Map the electrical connections in the power distribution room to a physical topology graph, use graph attention network to calculate the weights of adjacent device nodes under the sliding time window, and extract spatial association topology feature vectors.

[0008] S3. Calculate the Copula entropy between the environmental temperature and humidity data and the operating power data in the multimodal feature sequence, remove the temperature and humidity dependence in the spatial correlation topological feature vector based on the Copula entropy, and use a sparse Gaussian process to fit the decoupled spatial correlation topological feature vector to construct a dynamic energy consumption baseline with prediction variance.

[0009] S4. Compare the currently collected operating power data with the dynamic energy consumption baseline to mark abnormal nodes, convert the physical topology graph into an absorption state Markov chain transition matrix, calculate the absorption probability of walking to other device nodes with the abnormal node as the initial state, and output the probability topology tracing result.

[0010] S5. Using the probabilistic topological tracing results and the corresponding multimodal feature sequences as feedback samples, update the model parameters of the sparse Gaussian process and the network weights of the graph attention network.

[0011] Preferably, in step S1, the step of performing spatiotemporal alignment and interpolation preprocessing includes:

[0012] Extract the timestamp labels of the operating power data and the ambient temperature and humidity data, and use a cubic spline interpolation algorithm to fill in the missing sampling points;

[0013] Based on the step size of the sliding time window, the operating power data and the ambient temperature and humidity data with heterogeneous sampling frequencies are resampled to unify the time resolution.

[0014] The resampled operating power data and the environmental temperature and humidity data are subjected to zero-mean normalization to eliminate dimensional differences, and then spliced ​​together to generate the multimodal feature sequence.

[0015] Preferably, in step S2, the step of mapping the electrical connections within the substation to a physical topology diagram includes:

[0016] Each electrical device in the substation is defined as a discrete node, and the electrical transmission path between each electrical device is defined as a connection edge;

[0017] Extract the data slices of the multimodal feature sequence within the sliding time window as the initial feature matrix of the discrete node;

[0018] The admittance parameters and spatial distances between the discrete nodes are calculated by acquiring the distribution network structure data, and the static adjacency matrix of the connection edge is initialized accordingly.

[0019] Preferably, in step S2, the step of calculating the allocation weights of adjacent device nodes under the sliding time window includes:

[0020] A shared weight matrix is ​​used to perform a linear transformation on the data features of each device node in the multimodal feature sequence to obtain the latent feature vector of each device node;

[0021] The potential feature vectors of each device node and its corresponding neighboring device nodes are concatenated, and the scalar attention coefficient is calculated using a linear rectified function with leakage.

[0022] The scalar attention coefficients are normalized using a normalized exponential function, and the features of the adjacent device nodes are weighted and summed accordingly to output the spatial association topological feature vector.

[0023] Preferably, in step S3, the calculation step of the Copula entropy between the ambient temperature and humidity data and the operating power data includes:

[0024] Using a nonparametric kernel density estimation method, the marginal cumulative distribution functions of the environmental temperature and humidity data and the operating power data are calculated respectively.

[0025] Mapping the edge cumulative distribution function to a unit space, an empirical Copula density function reflecting the joint rank distribution relationship between the environmental temperature and humidity data and the operating power data is constructed.

[0026] The negative differential entropy value is obtained by performing a logarithmic integral operation on the empirical Copula density function.

[0027] Preferably, in step S3, the step of constructing a dynamic energy consumption baseline with prediction variance includes:

[0028] In the decoupled spatial correlation topological feature vector, a clustering algorithm is used to extract pseudo-input sample points as induced point vectors;

[0029] The covariance matrix between the spatially correlated topological feature vector and the induced point vector is calculated by selecting the squared exponential kernel function, thereby reducing the complexity of multidimensional distribution inversion;

[0030] Based on the variational inference principle, the lower bound of the logarithmic marginal likelihood is maximized, the expected value of the predicted distribution and the predicted variance are solved, and the dynamic energy consumption baseline is constructed.

[0031] Preferably, in step S4, the step of comparing the currently collected operating power data with the dynamic energy consumption baseline to mark abnormal nodes includes:

[0032] Extract the predicted expected value of the dynamic energy consumption baseline at the current acquisition time, and construct a confidence floating boundary by combining it with the prediction variance;

[0033] Substitute the currently collected operating power data into the confidence level floating boundary, and detect whether the fluctuation residual of the operating power data exceeds the boundary;

[0034] When the number of times the operating power data exceeds the confidence level floating boundary exceeds a preset threshold, the corresponding electrical equipment is marked as the abnormal node.

[0035] Preferably, in step S4, the step of converting the physical topology graph into an absorbing state Markov chain transition matrix includes:

[0036] Based on the device type, the device nodes in the physical topology diagram are divided into absorbing state nodes and transient state nodes;

[0037] The evolution probability between device nodes is calculated using the spatially associated topological feature vectors, and the absorbing state Markov chain transition matrix is ​​constructed.

[0038] Perform identity matrix inversion on the transient sub-blocks of the Markov chain transition matrix of the absorbing state to solve for the fundamental matrix, and calculate the absorption probability of walking to each of the absorbing state nodes.

[0039] Preferably, in step S5, the step of updating the model parameters of the sparse Gaussian process and the network weights of the graph attention network includes:

[0040] Construct a joint loss function that includes the root mean square error term for energy consumption prediction and the cross-entropy term for probabilistic source tracing to quantify the overall bias of the current model parameters;

[0041] Using an automatic differentiation framework and a backpropagation algorithm, the model gradients of the joint loss function with respect to the model parameters of the sparse Gaussian process and the network weights of the graph attention network are calculated simultaneously.

[0042] An optimizer with an adaptive momentum mechanism is used to perform parameter updates based on the model gradient.

[0043] The present invention has the following beneficial effects:

[0044] 1. In this invention, the interference of meteorological fluctuations is effectively removed by Copula entropy, and a dynamic energy consumption baseline with prediction variance is constructed by combining sparse Gaussian process. This breaks through the limitations of traditional static thresholds, reduces the computational complexity of high-dimensional data, and endows it with noise resistance stability. Thus, it can identify weak power consumption anomalies caused by the aging of real equipment with high confidence.

[0045] 2. In this invention, the spatial correlation features extracted by graph attention network are integrated into the absorbing state Markov chain transition matrix, realizing the leap from isolated equipment alarm to global probability tracing of the distribution network. This method can calculate the quantitative absorption probability of abnormal states traveling along electrical lines to various potential fault sources, providing a location reference for operation and maintenance personnel to investigate complex and hidden hidden dangers.

[0046] 3. In this invention, a joint loss function that integrates prediction error and source tracing cross-entropy is constructed. The source tracing results verified in the field are used as closed-loop samples to synchronously perform backpropagation and parameter updates. This enables the model to have the adaptive evolution capability to resist slow hardware aging and seasonal parameter drift, effectively avoiding the lag of offline retraining of traditional models and ensuring the long-term high-precision operation of the monitoring mechanism. Attached Figure Description

[0047] Figure 1 This is a flowchart of an AI-based method for identifying and optimizing energy consumption anomalies in power distribution substations, as proposed in this invention. Detailed Implementation

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] In embodiments of the present invention, the present invention provides an AI-based method for identifying and optimizing energy consumption anomalies in power distribution substations, such as... Figure 1 As shown, it includes:

[0050] S1. Establish a sliding time window, collect the operating power data and ambient temperature and humidity data of each electrical device in the substation, and perform spatiotemporal alignment and interpolation preprocessing to generate multimodal feature sequences.

[0051] Furthermore, step S1, the steps of performing spatiotemporal alignment and interpolation preprocessing, include:

[0052] Extract the timestamp labels from the operating power data and the ambient temperature and humidity data, and use a cubic spline interpolation algorithm to fill in the missing sampling points;

[0053] Based on the step size of the sliding time window, the operating power data and environmental temperature and humidity data with heterogeneous sampling frequencies are resampled to unify the time resolution.

[0054] Zero-mean normalization was performed on the resampled operating power data and the ambient temperature and humidity data to eliminate dimensional differences, and then the data were spliced ​​together to generate a multimodal feature sequence.

[0055] Specifically, to obtain high-quality underlying input features, the system first establishes a sliding time window based on a fixed time span and sliding step size. Within this window, real-time operating power data of each transformer and switchgear in the substation, as well as corresponding ambient temperature and humidity data, are continuously collected through smart meters and distributed environmental sensors. Due to differences in the sampling mechanisms of physical devices, after extracting the timestamp labels of the raw data stream, the system maps the operating power data and ambient temperature and humidity data onto a unified time axis for initial spatiotemporal alignment. In this scheme, the preferred time span of the sliding time window is 24 hours, and the preferred sliding step size is 15 minutes.

[0056] To address the issue of missing sampling points due to sensor communication delays or intermittent hardware failures, the system applies a cubic spline interpolation algorithm to numerically reconstruct the data sequence with extracted timestamp labels. Taking the running power data sequence as an example, let the normal collected power value at a known time node be... Within the interpolation interval formed by adjacent time points, the fitting function expression for the operating power is:

[0057] ;

[0058] in The target timestamp for the missing values ​​to be filled. This is the valid timestamp of the starting point of the interpolation interval. This represents the reconstructed operating power value at the target timestamp. (Variable) , , and For the spline interpolation polynomial coefficients, the system constructs a global linear equation system by simultaneously solving the physical boundary conditions of equal function values, continuous first derivatives, and continuous second derivatives at the connection points of adjacent piecewise curves. This allows for the analytical solution of the coefficients, thereby obtaining smooth, filler data that conforms to the actual changes in electricity load. The missing data for environmental temperature and humidity are filled using the same calculation logic. In a specific numerical implementation of this scheme, for a 15-minute missing data interval for a certain distribution cabinet, the coefficients of the polynomial constant term are obtained through analytical solution. The preferred value is 500, the coefficient of the first term. The preferred value is 2.5.

[0059] After filling in the missing sampling points, the system performs a resampling operation based on the sliding time window step size, taking into account the heterogeneous physical characteristics of operating power data (which is typically acquired at high frequencies) and ambient temperature and humidity data (which is typically acquired at low frequencies). The system regenerates a standard discrete time node sequence along the time axis at a uniform sliding step size interval, and substitutes these standard nodes into the constructed cubic spline interpolation function to calculate the corresponding values. This ensures that the data streams of all heterogeneous sensors in the substation are strictly aligned to the same sampling frequency and time resolution.

[0060] Because power data is measured in kilowatts and has large values, while temperature and humidity data are measured in degrees Celsius and percentages and have smaller values, the system performs zero-mean normalization on each dimension of the resampled data to eliminate the dimensional differences. The normalization formula is as follows:

[0061] ;

[0062] in Represents the first under a uniform time resolution Class of physical quantities in the first The actual values ​​at each time point, i.e., the corresponding operating power values ​​or ambient temperature and humidity values. The first [number]th ... The numerical mean of similar physical quantities The standard deviation of this type of data within the sliding time window. To standardize the eigenvalues ​​after eliminating dimensions, the system sequentially concatenates the standardized operating power, temperature, and humidity features extracted at the same standard time point, according to the time series, to generate a multimodal feature sequence matrix corresponding to the current sliding time window. In this scheme, the average value of the operating power is used for the transformer's rated operating state. The preferred power is 500 kilowatts, with a standard deviation of 100%. The preferred power is 50 kilowatts.

[0063] This step repaired the data gaps caused by abnormal transmission from the underlying IoT monitoring devices and smoothed out the differences in frequency and magnitude between multi-source heterogeneous sensor data, providing the model with high-quality front-end input features with a well-structured architecture.

[0064] S2. Map the electrical connections in the power distribution room to a physical topology graph, use graph attention network to calculate the weight distribution of adjacent device nodes under the sliding time window, and extract spatial correlation topology feature vectors.

[0065] Furthermore, step S2, which maps the electrical connections within the substation to a physical topology diagram, includes:

[0066] Each electrical device in the substation is defined as a discrete node, and the electrical transmission path between each electrical device is defined as a connection edge;

[0067] Extract data slices of multimodal feature sequences within a sliding time window as the initial feature matrix for discrete nodes;

[0068] Obtain the distribution network structure data, calculate the admittance parameters and spatial distances between discrete nodes, and initialize the static adjacency matrix of the connection edge accordingly.

[0069] Further, step S2, the step of calculating the allocation weights of adjacent device nodes under the sliding time window, includes:

[0070] By using a shared weight matrix to perform a linear transformation on the data features of each device node in the multimodal feature sequence, the latent feature vectors of each device node can be obtained.

[0071] The potential feature vectors of each device node and its corresponding neighboring device nodes are concatenated, and the scalar attention coefficient is calculated using a linear rectified function with leakage.

[0072] The scalar attention coefficients are normalized using a normalized exponential function, and the features of adjacent device nodes are weighted and summed accordingly to output a spatial correlation topological feature vector.

[0073] Specifically, the system first analyzes the internal wiring layout and spatial plan of the substation, defining the actual physical electrical equipment such as transformers, switchgear at all levels, and reactive power compensation devices as discrete nodes in the graph structure. Simultaneously, the system defines the actual physical electrical transmission paths, such as cables and busbars laid between the equipment, as connection edges, thereby completing the abstract mapping of the distribution network topology. For each constructed discrete node, the system extracts the corresponding standardized operating power data and environmental temperature and humidity data slices from the preprocessing results output in step S1, according to a given sliding time window length. These time-series multidimensional data slices are combined and flattened to construct an initial feature matrix representing the initial operating state of the corresponding discrete equipment node. Subsequently, the system obtains the cable material properties and laying trajectory lengths from the distribution network structure data, and calculates the electrical admittance parameters and physical spatial distances between adjacent discrete nodes. The system uses the calculated electrical admittance parameter amplitude and the reciprocal of the physical spatial distance to perform a weighted summation to calculate the connectivity strength, and uses this as a basis to initialize the static adjacency matrix corresponding to all connection edges, determining the basic spatial topological connectivity domain between each equipment node.

[0074] After defining the physical topology graph, the system introduces a graph attention network to compute the feature interactions between nodes. In this specific implementation, the graph attention network is configured with multiple layers of graph attention hidden layers, preferably three layers. Each hidden layer contains a linear layer with 64 neurons to unify the dimensionality of the high-dimensional feature space, and the single-layer feedforward neural network used to compute feature dependency strength is configured as a scalar output structure with one output node. The system sets a shared weight matrix containing learnable weights to perform high-dimensional mapping on the initial feature matrices of each electrical device node in the multimodal feature sequence. The mathematical expression for this linear transformation process is:

[0075] ;

[0076] in Indicates the first The initial data feature vector of each electrical equipment node, that is, the input feature vector containing the node's operating power and ambient temperature and humidity. This represents a shared weight matrix used by all nodes. This is the latent feature vector extracted from the electrical equipment node after feature mapping processing, representing its high-dimensional power consumption behavior characteristics. In this scheme, a shared weight matrix is ​​used. The output dimension is preferably 64 dimensions.

[0077] After obtaining the latent feature vectors of all device nodes, for any target device node and its adjacent device nodes defined by the static adjacency matrix in the physical topology graph, the system concatenates the latent feature vectors of the target device node and its corresponding adjacent device nodes in dimensional space. The system then inputs the concatenated vector into the configured single-layer feedforward neural network and calculates the feature interdependence strength between devices using a leaky linear rectified function. The formula for calculating this scalar attention coefficient is as follows:

[0078] ;

[0079] in For target device node latent feature vectors Adjacent device nodes latent feature vectors This represents the concatenation operation of two high-dimensional vectors. This represents the transposed weight vector in a single-layer feedforward neural network. For a linear rectified activation function with leakage, For the target device node of the calculation output With adjacent device nodes The scalar attention coefficient between adjacent nodes, in a physical sense, preliminarily quantifies the expected impact of power load fluctuations at neighboring nodes on the state of the central target node. In this scheme, the leakage-triggered linear rectification activation function... The preferred slope parameter for the negative half-axis is 0.2.

[0080] To ensure that the influence weights of different neighboring nodes in the local topology network conform to a probability distribution and are numerically comparable, the system uses a normalized exponential function to normalize all scalar attention coefficients within the neighborhood of nodes belonging to the same target device. The distribution calculation formula is as follows:

[0081] ;

[0082] in Represents the target device node defined by the static adjacency matrix of the distribution network. The set of all electrically adjacent nodes, Exponentiation of the base of the natural constant. To calculate and obtain adjacent device nodes Relative to the target device node The final weight allocation within the current sliding time window. Based on the weights allocated to all neighboring nodes calculated above, the system performs a weighted summation and aggregation of the latent feature vectors of neighboring device nodes. The mathematical expression for the aggregated output is as follows:

[0083] ;

[0084] in Represents a nonlinear activation function, calculated as follows: This refers to a spatially correlated topological feature vector that integrates the electrical connectivity topology and the collaborative operational status information of surrounding equipment. In this scheme, a nonlinear activation function... The preferred option is the ELU function.

[0085] This step transforms isolated equipment monitoring data into graph topology data that includes the physical connectivity of electrical lines, enabling the system to perceive the cross-node coupling propagation patterns of energy consumption anomalies between local power distribution equipment.

[0086] S3. Calculate the Copula entropy between environmental temperature and humidity data and operating power data in the multimodal feature sequence. Based on the Copula entropy, remove the temperature and humidity dependence in the spatial correlation topological feature vector. Use a sparse Gaussian process to fit the decoupled spatial correlation topological feature vector and construct a dynamic energy consumption baseline with prediction variance.

[0087] Furthermore, step S3, the calculation of the Copula entropy between the ambient temperature and humidity data and the operating power data, includes:

[0088] The marginal cumulative distribution functions of environmental temperature and humidity data and operating power data were calculated using a nonparametric kernel density estimation method.

[0089] By mapping the marginal cumulative distribution function to a unit space, an empirical Copula density function is constructed that reflects the joint rank distribution relationship between environmental temperature and humidity data and operating power data.

[0090] The Copula entropy is obtained by performing a logarithmic integral on the empirical Copula density function to solve for its negative differential entropy.

[0091] Furthermore, step S3, the step of constructing a dynamic energy consumption baseline with prediction variance, includes:

[0092] In the decoupled spatial correlation topological feature vector, a clustering algorithm is used to extract pseudo-input sample points as induced point vectors;

[0093] The squared exponential kernel function is selected to calculate the covariance matrix between the spatial correlation topological feature vector and the induced point vector, thereby reducing the complexity of multidimensional distribution inversion.

[0094] Based on the variational inference principle, the lower bound of the logarithmic marginal likelihood is maximized to solve for the expected value and variance of the prediction distribution, and a dynamic energy consumption baseline is constructed.

[0095] Specifically, a kernel density estimation model is used to process multimodal feature sequences. For environmental and power consumption variables, the system calculates the marginal cumulative distribution functions for environmental temperature and humidity data and operating power data, respectively. The formula for obtaining the marginal cumulative distribution probability by integrating the kernel density function is as follows:

[0096] ;

[0097] in This represents the environmental temperature and humidity observation sample value or the operating power observation sample value within the current time window. The total number of observed samples, For smoothing kernel function, The bandwidth parameter is used to control the smoothness. This is used to calculate the marginal cumulative distribution probability values ​​of the output ambient temperature and humidity or operating power. In this scheme, a smoothing kernel function is used. The preferred kernel function is a Gaussian kernel, and the bandwidth parameter controls the smoothness. The preferred value is 0.05, representing the total number of observation samples extracted within the sliding time window. The preferred value is 96.

[0098] The system maps the marginal cumulative distribution probability domains of the obtained physical parameters to a unit space between zero and one, thereby constructing an empirical Copula density function reflecting the joint rank distribution relationship between environmental temperature and humidity data and operating power data. Logarithmic integration is performed on this density function to solve for the negative differential entropy value. The discretized evaluation formula for calculating the Copula entropy is as follows:

[0099] ;

[0100] in and These represent the edge cumulative probability distribution values ​​of the synchronously sampled ambient temperature and humidity data and operating power data after mapping to a unit space, respectively. Represents the empirical Copula density function value. To obtain the Copula entropy, a value that precisely quantifies the nonlinear interference intensity of meteorological and environmental fluctuations on the actual power consumption of equipment, the system constructs feature penalty weights using the calculated Copula entropy. A linear projection transformation is then performed on the spatially correlated topological feature vector, and data dimensions highly correlated with temperature and humidity are filtered out according to the penalty weight ratio. This process yields a decoupled spatially correlated topological feature vector free from environmental interference.

[0101] K-means clustering is applied to the decoupled spatially correlated topological feature vectors to divide the massive high-dimensional feature vector set into significantly fewer clusters. The geometric centroids of each cluster are extracted as pseudo-input sample points to form induced point vectors. The system selects the squared exponential kernel function to quantify the nonlinear similarity in the multidimensional data space, calculating the cross-domain covariance matrix between the decoupled spatially correlated topological feature vectors and the induced point vectors, as well as the internal autocovariance matrix of the induced point vectors. The formula for calculating the squared exponential kernel function is as follows:

[0102] ;

[0103] in This represents the spatial correlation topological feature vector after decoupling. This represents the extracted induced point vector. The signal variance is used to control the fluctuation amplitude of the function. To control the characteristic length scale of correlation decay rate, This is the output kernel function covariance element representing the spatial correlation of the data. This calculation process reduces the complex operation of inverting the covariance matrix of all physical device nodes to a low-order matrix dimension composed of a finite number of induced points. In this scheme, the signal variance... Preferably 1, feature length scale The preferred value is 0.5, and the preferred number of induced points for pseudo-input sample points is 50.

[0104] Based on the principle of variational inference, a lower bound for evidence using logarithmic marginal likelihood is constructed. A gradient optimization algorithm is then used to continuously update the hyperparameters of the induced point distribution and the kernel function to maximize this lower bound. After optimization convergence, the system calculates the expected value and variance of the predicted distribution of the power distribution equipment's operating state based on the current decoupling characteristics. The dynamic prediction formula is as follows:

[0105] ;

[0106] ;

[0107] in Let be the covariance matrix between the feature vector of the device to be predicted and the induced point vector. Let be the autocovariance matrix inside the induced point vector. It is the transpose of the cross-domain covariance matrix. Let be the covariance matrix of the feature vector of the device to be predicted. and These are the mean vector and covariance matrix of the induced points obtained from the variational distribution optimization solution, respectively. This represents the expected power consumption under standard operating conditions predicted by the system. This represents the prediction variance output by the prediction model. The system combines the expected standard operating power consumption at continuous timestamps with its prediction variance to construct a dynamic energy consumption baseline that adaptively adjusts the fluctuation range according to actual operating conditions.

[0108] This step accurately eliminated normal load fluctuations caused by weather changes through multidimensional information entropy calculation, and significantly reduced the computational power consumption of predicting the topological features of massive nodes by using a sparse Bayesian model that introduces induction points, providing a benchmark reference with mathematical confidence boundaries for subsequent anomaly comparison.

[0109] S4. Compare the currently collected operating power data with the dynamic energy consumption baseline to mark abnormal nodes, transform the physical topology graph into an absorption state Markov chain transition matrix, calculate the absorption probability of walking to other device nodes with the abnormal nodes as the initial state, and output the probabilistic topology tracing results.

[0110] Further, step S4, which involves comparing the currently collected operating power data with the dynamic energy consumption baseline to mark abnormal nodes, includes:

[0111] Extract the expected value of the dynamic energy consumption baseline at the current acquisition time, and construct the confidence floating boundary by combining it with the prediction variance;

[0112] Substitute the currently collected operating power data into the confidence level floating boundary and check whether the fluctuation residual of the operating power data exceeds the boundary.

[0113] When the number of times the operating power data exceeds the confidence level floating boundary exceeds a preset threshold, the corresponding electrical equipment will be marked as an abnormal node.

[0114] Further, step S4, the step of transforming the physical topology graph into an absorbing state Markov chain transition matrix, includes:

[0115] Based on the device type, the device nodes in the physical topology diagram are divided into absorbing state nodes and transient state nodes;

[0116] The evolution probability between device nodes is calculated using spatially correlated topological feature vectors, and the transition matrix of the absorbing state Markov chain is constructed.

[0117] Perform identity matrix inversion on the transient sub-blocks of the Markov chain transition matrix of the absorbing state to solve for the fundamental matrix, and calculate the absorption probability of walking to each absorbing state node.

[0118] Specifically, the system extracts the expected prediction value and prediction variance at the current data acquisition moment from the self-constructed dynamic energy consumption baseline. The system takes the square root of the prediction variance to obtain the prediction standard deviation, sets a confidence coefficient to adjust the sensitivity, and then adds or subtracts the product of the prediction standard deviation and the confidence coefficient from the expected prediction value to construct a confidence level floating boundary consisting of an upper and lower prediction limit. The formula for calculating the confidence level floating boundary is as follows:

[0119] ;

[0120] ;

[0121] in This is the current data collection time. This is the expected value of the extracted dynamic energy consumption baseline at the current moment. The standard deviation of the forecast at the current moment. Here is the confidence coefficient. For the upper limit of the prediction, The system uses the currently collected operating power data as the prediction lower bound. It substitutes this confidence level floating boundary to determine if the actual value lies between the prediction upper and lower bounds. If the operating power data is greater than the prediction upper bound or less than the prediction lower bound, the system detects that the residual fluctuation of the current operating power data exceeds this boundary. The system is configured with a timing-tolerant counter. When the number of times the operating power data of a specific electrical device exceeds the confidence level floating boundary exceeds a preset threshold, it determines that the device has persistent power consumption distortion and marks the corresponding electrical device as an abnormal node. In this scheme, the confidence coefficient of the sensitivity is adjusted. Preferably, the value is 3.0, and the preset threshold for timing fault tolerance is preferably 5 times.

[0122] After locating the abnormal nodes, the system classifies the equipment nodes in the physical topology map into absorptive nodes and transient nodes based on the physical power transmission and distribution attributes of the equipment types within the substation. Terminal load equipment or distribution transformers representing the energy transmission endpoints in the distribution network are designated as absorptive nodes, while switchgear and busbars representing intermediate energy convergence or branch transmission paths are designated as transient nodes. The system uses the previously extracted spatial correlation topology feature vectors to calculate the electrical connectivity similarity between equipment nodes, thereby characterizing the evolution probability of abnormal states among the nodes. The formula for calculating the evolution probability is as follows:

[0123] ;

[0124] in For target device node Spatial correlation topological feature vectors and Adjacent device nodes and Spatial correlation topological feature vectors The characteristic transformation matrix, Represents the target device node The set of all adjacent nodes, For exponential operations, To calculate the output nodes To the node The evolution probabilities of the transitions. The system fills all the calculated evolution probability elements into a square matrix to construct a structured absorbing Markov chain transition matrix. In this scheme, the characteristic transformation matrix... The preferred dimension configuration is 64×64.

[0125] The system structurally divides the transition matrix of the absorbing Markov chain into sub-blocks, extracting transient sub-blocks representing the probability of transitions between transient nodes and absorbing transition sub-blocks representing the probability of transitions from transient nodes to absorbing nodes. To quantify the global expectation of the topology walk path, the system constructs an identity matrix with the same dimension as the transient sub-blocks. The transient sub-blocks are subtracted from the identity matrix, and the resulting matrix blocks are inverted to solve for the fundamental matrix. The analytical formula for the fundamental matrix is ​​as follows:

[0126] ;

[0127] in This refers to the transient sub-block extracted from the transition matrix of the absorbing Markov chain. It is the identity matrix. To solve for the Markov chain's fundamental matrix, whose internal elements represent the expected number of stops during anomaly state walks between transient nodes, the system then multiplies the fundamental matrix by the absorbing transition sub-block to calculate the absorption probability of a walk from any transient node converging to any absorbing state node. The formula for calculating the absorption probability is as follows:

[0128] ;

[0129] in To characterize the absorption transfer sub-block that transitions from a transient node to an absorbing node, This is the output absorption probability matrix. The system takes the marked abnormal node as the initial state, extracts the row vector corresponding to the abnormal node from the absorption probability matrix, and outputs the exact probability distribution of the fault source of the abnormal node pointing to different absorption state nodes, as the final probabilistic topological tracing result.

[0130] This step transforms the system from simple threshold over-limit alarms to root cause localization of the distribution network's physical topology, providing quantitative statistical data support for troubleshooting fault sources.

[0131] S5. Use the probabilistic topological tracing results and the corresponding multimodal feature sequences as feedback samples to update the model parameters of the sparse Gaussian process and the network weights of the graph attention network.

[0132] Furthermore, step S5, which involves updating the model parameters of the sparse Gaussian process and the network weights of the graph attention network, includes:

[0133] Construct a joint loss function that includes the root mean square error term for energy consumption prediction and the cross-entropy term for probabilistic source tracing to quantify the overall bias of the current model parameters;

[0134] Using an automatic differentiation framework and a backpropagation algorithm, we simultaneously calculate the model gradients of the joint loss function on the model parameters of the sparse Gaussian process and the network weights of the graph attention network.

[0135] An optimizer with an adaptive momentum mechanism is used to perform parameter updates based on model gradients.

[0136] Specifically, during operation, real fault records verified by on-site personnel or automated systems are continuously accumulated. The system extracts and combines the historically accumulated probabilistic topological tracing results and corresponding multimodal feature sequences as a complete feedback sample set input into the training process. To accurately quantify the deviation between the model output and the actual situation, the system constructs a joint loss function including a root mean square error term for energy consumption prediction and a cross-entropy term for probabilistic tracing. The root mean square error term for energy consumption prediction measures the deviation between the prediction baseline constructed by the sparse Gaussian process and the actual power consumption of the device; its calculation formula is as follows:

[0137] ;

[0138] in This represents the total number of samples in the current feedback training batch. The model predicts the first [number]th ... The expected standard operating power consumption of a sample The observed value is the actual operating power corresponding to this sample. The probabilistic source tracing cross-entropy term is used to measure the statistical difference between the source tracing probability distribution generated by the graph attention network and the actual fault node distribution. Its calculation formula is as follows:

[0139] ;

[0140] in This represents the total number of absorbing state nodes defined in the physical topology graph. For a true fault indication label, when the first The actual source of the fault in the sample was identified as the first... The value is one when there are absorbing nodes, and zero otherwise. To calculate the output of the first The sample traveled to the first... The system predicts the absorption probability of each absorbing node. It constructs a joint loss function by linearly weighting the regression loss and classification loss using a balance coefficient.

[0141] ;

[0142] in and These are the weighting coefficients that control the gradient contribution of the regression task and the source classification task, respectively. To quantify the overall evaluation value of the current model parameter bias, in this scheme, the total number of samples in the current feedback training batch is used. The optimal value is 128, which controls the weighting coefficients of the regression task. The optimal value is 0.7, which controls the weighting coefficient of the source tracing and classification task. The preferred value is 0.3.

[0143] The system deploys the constructed joint loss function in a computational graph engine that supports automatic differentiation. During the backpropagation computation phase, the system invokes the automatic differentiation framework based on the chain rule to synchronously and accurately calculate the partial derivatives of the joint loss function with respect to each parameter to be optimized at the lower levels. For the sparse Gaussian process part, the system calculates the gradient vector of the loss value with respect to the feature position matrix of the induced points and the length scaling parameter of the kernel function. For the graph attention network part, the system synchronously calculates the gradient matrix of the loss value with respect to the shared feature transformation matrix and the scalar attention weights of each layer. The mathematically unified definition of the model gradient is...

[0144] ;

[0145] in This represents the set of all learnable model parameter vectors at the current iteration step, including the weights of the sparse Gaussian process and the underlying network of the graph attention network. The calculated set of model parameter gradients indicates the fastest parameter adjustment direction to reduce diagnostic errors.

[0146] After extracting the model gradients, the system calls an optimizer algorithm with adaptive momentum to update the parameters. In each training batch iteration, the system updates the momentum variable and the adaptive learning rate parameter using the currently extracted gradient values. The update formula for the momentum variable is as follows:

[0147] ;

[0148] ;

[0149] in This is the cumulative value of the first-order momentum variable. This is the cumulative value of the second-order momentum variable. and These are the exponentially decaying hyperparameters controlling the retention rates of historical information for the first and second moments, respectively. After correcting for initial bias in the calculated momentum variables, the system performs specific parameter update actions based on the corrected model gradient. The update calculation formula is as follows:

[0150] ;

[0151] in This is the set of parameter vectors for the newly updated model. The system's preset base learning rate parameter, and These are the first-order and second-order momentum predictions after bias correction. To prevent numerical smoothing constants with zero denominators, the system iteratively executes forward inference and backward parameter update processes for batch data until the overall evaluation value converges. In this scheme, the exponential decay hyperparameter of the first-moment estimation is controlled. The optimal value is 0.9, which controls the exponential decay hyperparameter of the second-order moment estimation. The preferred value is 0.999, which is the system's preset base learning rate parameter. The preferred value is 0.001, a numerical smoothing constant to prevent the denominator from being zero in calculations. The preferred value is 1e-8.

[0152] This step enables the anomaly identification model to continuously self-evolve, automatically calibrating the prediction baseline and optimizing the topology tracing path in response to seasonal environmental parameter drift and the slow aging of power distribution station hardware.

[0153] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms, characterized in that, include: S1. Establish a sliding time window, collect the operating power data and ambient temperature and humidity data of each electrical device in the substation, and perform spatiotemporal alignment and interpolation preprocessing to generate multimodal feature sequences. S2. Map the electrical connections in the power distribution room to a physical topology graph, use graph attention network to calculate the weights of adjacent device nodes under the sliding time window, and extract spatial association topology feature vectors. S3. Calculate the Copula entropy between the environmental temperature and humidity data and the operating power data in the multimodal feature sequence, remove the temperature and humidity dependence in the spatial correlation topological feature vector based on the Copula entropy, and use a sparse Gaussian process to fit the decoupled spatial correlation topological feature vector to construct a dynamic energy consumption baseline with prediction variance. S4. Compare the currently collected operating power data with the dynamic energy consumption baseline to mark abnormal nodes, convert the physical topology graph into an absorption state Markov chain transition matrix, calculate the absorption probability of walking to other device nodes with the abnormal node as the initial state, and output the probability topology tracing result. S5. Using the probabilistic topological tracing results and the corresponding multimodal feature sequences as feedback samples, update the model parameters of the sparse Gaussian process and the network weights of the graph attention network.

2. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S1, the step of performing spatiotemporal alignment and interpolation preprocessing, includes: Extract the timestamp labels of the operating power data and the ambient temperature and humidity data, and use a cubic spline interpolation algorithm to fill in the missing sampling points; Based on the step size of the sliding time window, the operating power data and the ambient temperature and humidity data with heterogeneous sampling frequencies are resampled to unify the time resolution. The resampled operating power data and the environmental temperature and humidity data are subjected to zero-mean normalization to eliminate dimensional differences, and then spliced ​​together to generate the multimodal feature sequence.

3. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S2, the step of mapping the electrical connections within the substation to a physical topology diagram, includes: Each electrical device in the substation is defined as a discrete node, and the electrical transmission path between each electrical device is defined as a connection edge; Extract the data slices of the multimodal feature sequence within the sliding time window as the initial feature matrix of the discrete node; The admittance parameters and spatial distances between the discrete nodes are calculated by acquiring the distribution network structure data, and the static adjacency matrix of the connection edge is initialized accordingly.

4. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S2, the step of calculating the allocation weights of adjacent device nodes under the sliding time window, includes: A shared weight matrix is ​​used to perform a linear transformation on the data features of each device node in the multimodal feature sequence to obtain the latent feature vector of each device node; The potential feature vectors of each device node and its corresponding neighboring device nodes are concatenated, and the scalar attention coefficient is calculated using a linear rectified function with leakage. The scalar attention coefficients are normalized using a normalized exponential function, and the features of the adjacent device nodes are weighted and summed accordingly to output the spatial association topological feature vector.

5. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S3, the calculation of the Copula entropy between the ambient temperature and humidity data and the operating power data, includes: Using a nonparametric kernel density estimation method, the marginal cumulative distribution functions of the environmental temperature and humidity data and the operating power data are calculated respectively. Mapping the edge cumulative distribution function to a unit space, an empirical Copula density function reflecting the joint rank distribution relationship between the environmental temperature and humidity data and the operating power data is constructed. The negative differential entropy value is obtained by performing a logarithmic integral operation on the empirical Copula density function.

6. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S3, the step of constructing a dynamic energy consumption baseline with prediction variance, includes: In the decoupled spatial correlation topological feature vector, a clustering algorithm is used to extract pseudo-input sample points as induced point vectors; The covariance matrix between the spatially correlated topological feature vector and the induced point vector is calculated by selecting the squared exponential kernel function, thereby reducing the complexity of multidimensional distribution inversion; Based on the variational inference principle, the lower bound of the logarithmic marginal likelihood is maximized, the expected value of the predicted distribution and the predicted variance are solved, and the dynamic energy consumption baseline is constructed.

7. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S4, the step of comparing the currently collected operating power data with the dynamic energy consumption baseline to mark abnormal nodes, includes: Extract the predicted expected value of the dynamic energy consumption baseline at the current acquisition time, and construct a confidence floating boundary by combining it with the prediction variance; Substitute the currently collected operating power data into the confidence level floating boundary, and detect whether the fluctuation residual of the operating power data exceeds the boundary; When the number of times the operating power data exceeds the confidence level floating boundary exceeds a preset threshold, the corresponding electrical equipment is marked as the abnormal node.

8. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S4, the step of converting the physical topology graph into an absorbing state Markov chain transition matrix, includes: Based on the device type, the device nodes in the physical topology diagram are divided into absorbing state nodes and transient state nodes; The evolution probability between device nodes is calculated using the spatially associated topological feature vectors, and the absorbing state Markov chain transition matrix is ​​constructed. Perform identity matrix inversion on the transient sub-blocks of the Markov chain transition matrix of the absorbing state to solve for the fundamental matrix, and calculate the absorption probability of walking to each of the absorbing state nodes.

9. The method for identifying and optimizing energy consumption anomalies in substations based on AI algorithms according to claim 1, characterized in that, Step S5, the step of updating the model parameters of the sparse Gaussian process and the network weights of the graph attention network, includes: Construct a joint loss function that includes the root mean square error term for energy consumption prediction and the cross-entropy term for probabilistic source tracing to quantify the overall bias of the current model parameters; Using an automatic differentiation framework and a backpropagation algorithm, the model gradients of the joint loss function with respect to the model parameters of the sparse Gaussian process and the network weights of the graph attention network are calculated simultaneously. An optimizer with an adaptive momentum mechanism is used to perform parameter updates based on the model gradient.

Images