Real-time energy consumption abnormal data monitoring method of space-time object multivariate fusion model

Abstract

The invention discloses a real-time energy consumption abnormal data monitoring method of a space-time object multivariate fusion model, which belongs to the technical field of energy consumption monitoring and comprises the following steps: S1, constructing a data set; s2, constructing a training space-time object multivariate fusion model; s3, collecting energy consumption data in real time; s4, predicting whether abnormal data exists in the real-time energy consumption data or not; s5, predicting whether abnormal data exist in the real-time energy consumption data by the improved Pauta method; according to the method, time, space, environment and equipment attributes acted by real-time energy consumption can be considered at the same time, and the prediction precision is improved; and meanwhile, the problem of monitoring errors caused by relatively large numerical value change of the energy consumption data along with the startup and shutdown of the equipment at a certain time is avoided, and the prediction precision is further improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of energy consumption monitoring, and particularly relates to a real-time energy consumption abnormal data monitoring method of a space-time object multi-element fusion model. BACKGROUND

[0002] With the gradual development of the "double carbon" process, ensuring a safe and efficient modern energy system is an inherent requirement for ensuring national energy security and striving to achieve carbon peak and carbon neutrality on schedule, and is an important support for promoting high-quality economic and social development.

[0003] Real-time energy consumption abnormal data identification and monitoring is one of the important means of ensuring a safe and efficient modern energy system, which can timely discover system problems and ensure safe and stable operation of the system.

[0004] At present, the real-time energy consumption abnormal data identification and monitoring method generally trains a time series model through historical time series data for prediction to identify and monitor real-time energy consumption abnormal data. This method does not take into account the characteristics of the space and energy consumption generating equipment affected by real-time energy consumption, which can easily affect the prediction result and reduce the prediction accuracy.

[0005] In view of this, a real-time energy consumption abnormal data monitoring method of a space-time object multi-element fusion model is designed to solve the above problems. SUMMARY

[0006] To solve the problems proposed in the background, the application provides a real-time energy consumption abnormal data monitoring method of a space-time object multi-element fusion model, which has the characteristics of improving prediction accuracy.

[0007] To achieve the above purpose, the application provides the following technical scheme: a real-time energy consumption abnormal data monitoring method of a space-time object multi-element fusion model, comprising the following steps:

[0008] S1: collecting energy consumption historical data and pre-processing, saving the pre-processed energy consumption historical data to a database as a data set;

[0009] S2: constructing a space-time object multi-element fusion model, training through the data set until convergence;

[0010] S3: real-time collection of energy consumption data;

[0011] S4: predicting whether there is abnormal data in the real-time energy consumption data based on the converged space-time object multi-element fusion model;

[0012] S5: input the real-time energy consumption data and the spatiotemporal object multi-fusion model prediction result to the improved Rayleigh method, collect the energy consumption data within 24 hours at the same time, calculate the standard deviation and the mean value, based on the real-time energy consumption data, the spatiotemporal object multi-fusion model prediction result, the mean value of the energy consumption data within 24 hours and the self-defined parameter, calculate the residual error, if the deviation value is less than three times the standard deviation of the energy consumption data within 24 hours, it is an abnormal value, otherwise, it is a normal value, and output the identification result of the real-time energy consumption data.

[0013] Further, in the step S1, the energy consumption historical data includes time, space size, environment, device attribute and energy consumption value dimensions.

[0014] Further, in the step S1, the energy consumption historical data preprocessing includes the following steps:

[0015] The collected energy consumption historical data is cleaned to remove abnormal data noise, and missing data is supplemented through spline interpolation processing;

[0016] The energy consumption historical data is uniformly represented according to international standard units;

[0017] The energy consumption historical data is normalized according to the maximum energy consumption calculation result;

[0018] The features of the energy consumption historical data are extracted and represented as fixed-size vectors;

[0019] The fixed-size vector-represented energy consumption historical data features are constructed into a directed acyclic graph according to time sequence, the nodes in the graph represent the energy consumption data features at a certain time, and the edges represent the dependency relationship between the nodes, i.e. the node features.

[0020] Further, in the step S2, the constructed spatiotemporal object multi-fusion model includes a GNN-Transformer model, the GNN-Transformer model includes a GNN model, a Transformer model and a classifier, the GNN model includes a time GNN model, a space GNN model, an environment GNN model and a device GNN model, and the time GNN model includes feature splicing.

[0021] The time and energy consumption value dimensions are encoded as the node features of the time GNN model, and the t+1 time update expression of the time GNN model is:

[0022]

[0023] In the formula: f represents a state update function of a hidden state, which is a global shared function; x e o[v] represents the features of the edges adjacent to the node v; is the hidden state of the adjacent node at time t; x ne[v] represents the features of the neighbor nodes of node v;

[0024] The energy consumption value dimension is set by the spatial feature, the spatial size and the energy consumption value dimension are encoded, and the spatial GNN model is used as the node feature of the spatial GNN model.

[0025] The energy consumption value dimension is set by the environmental factor feature, the environmental data and the energy consumption value dimension are encoded, and the environmental GNN model is used as the node feature of the environmental GNN model.

[0026] The energy consumption value dimension is set by the device attribute feature, the device attribute data and the energy consumption value dimension are encoded, and the device GNN model is used as the node feature of the device GNN model.

[0027] The attention weight expression of the Transformer model is:

[0028]

[0029] In the formula, Q represents a query matrix, K represents a key matrix, V represents a value matrix, d k The dimension of the key is used to scale to avoid gradient disappearance caused by too large dot product.

[0030] Further, in the step S2, the data set is divided into a test set and a validation set, the test set is used for testing, the hyperparameters are adjusted, and the performance and accuracy are evaluated, and the validation set is used for verification.

[0031] Further, the specific steps of the step S4 include:

[0032] The time GNN model is based on the time and energy consumption value dimensions, captures the local dynamic dependency relationship between nodes, and generates node representation containing structure information, i.e., node embedding.

[0033] The spatial GNN model is based on the spatial size and energy consumption value dimensions, captures the local dynamic dependency relationship between nodes, and generates node representation containing structure information, i.e., node embedding.

[0034] The environmental GNN model is based on the environmental data and energy consumption value dimensions, captures the local dynamic dependency relationship between nodes, and generates node representation containing structure information, i.e., node embedding.

[0035] The device GNN model is based on the device attribute data and energy consumption value dimensions, captures the local dynamic dependency relationship between nodes, and generates node representation containing structure information, i.e., node embedding.

[0036] The node embeddings of the time GNN model, the spatial GNN model, the environmental GNN model and the device GNN model are spliced and input into the Transformer model.

[0037] The Transformer model captures long-distance dependencies across nodes through a self-attention mechanism, dynamically assigns weights to the association of different nodes, and concatenates local feature dependencies into global feature dependencies;

[0038] The linear transformation layer maps the global features output by the Transformer model to the required result dimension;

[0039] The Softmax layer converts the linear output into a probability distribution and outputs the prediction result based on a preset decision rule.

[0040] Further, in the step S5, the expression of the improved Ludada method is:

[0041]

[0042] In the formula, x represents real-time energy consumption data; represents the mean value of real-time energy consumption data within 24 hours; x forecast represents the predicted value, the weight of the predicted value is increased; a represents a custom parameter; s represents the standard deviation of real-time energy consumption data within 24 hours;

[0043]

[0044] In the formula, n represents the number of samples; represents the deviation between a certain data and the sample mean.

[0045] Compared with the prior art, the beneficial effects of the present application are:

[0046] The GNN model and the Transformer model are used to learn the characteristics of short-time data and long-time data of energy consumption in time, space, environment and equipment properties, the model based on the characteristic learning monitors the abnormal data in real-time energy consumption data, can consider the time, space, environment and equipment properties affected by real-time energy consumption at the same time, improve the prediction accuracy; the prediction result output by the model increases the weight of the predicted value through the improved Ludada method, avoids the monitoring error problem caused by the large numerical change of energy consumption data at a certain time with the on-off of the equipment, and further improves the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The method flowchart of the present application;

[0048] Figure 2 The energy consumption historical data preprocessing flowchart of the present application;

[0049] Figure 3 The model processing flowchart of the present application;

[0050] Figure 4 Flow chart of improved Relyda method for identifying abnormal data. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0052] The present application provides the following technical solutions: a real-time energy consumption abnormal data monitoring method of a space-time object multi-element fusion model, comprising the following steps:

[0053] S1: collecting energy consumption historical data and performing preprocessing, saving the preprocessed energy consumption historical data into a database as a data set;

[0054] S2: constructing a space-time object multi-element fusion model, training through the data set until convergence;

[0055] S3: collecting real-time energy consumption data;

[0056] S4: predicting whether there is abnormal data in the real-time energy consumption data based on the converged space-time object multi-element fusion model;

[0057] S5: inputting the real-time energy consumption data and the space-time object multi-element fusion model prediction result to an improved Relyda method, collecting energy consumption data within 24 hours, calculating the standard deviation and the mean value, calculating the residual based on the real-time energy consumption data, the space-time object multi-element fusion model prediction result, the energy consumption data mean value within 24 hours and the self-defined parameter, if the deviation value is less than three times the standard deviation of the energy consumption data within 24 hours, it is an abnormal value, otherwise, it is a normal value, and outputting the identification result of the real-time energy consumption data.

[0058] Specifically, in step S1, the energy consumption historical data includes time, space size, environment, device attribute and energy consumption value dimensions.

[0059] Specifically, in step S1, the energy consumption historical data preprocessing comprises the following steps:

[0060] The collected energy consumption historical data is cleaned to remove abnormal data noise, and missing data is supplemented through spline interpolation processing;

[0061] The energy consumption historical data is uniformly represented according to international standard units;

[0062] The energy consumption historical data is calculated according to international standard units, which can avoid the calculation error caused by the non-uniformity of units;

[0063] The energy consumption history data is normalized according to the maximum energy consumption calculation result;

[0064] The energy consumption history data is normalized according to the maximum energy consumption calculation result, which can unify the data feature analysis and measurement scale;

[0065] The features of the energy consumption history data are extracted and represented as fixed-size vectors;

[0066] The energy consumption history data is represented as fixed-size vectors, which can unify the data for processing;

[0067] The energy consumption history data features represented by fixed-size vectors are constructed into a directed acyclic graph according to the time sequence, and the nodes in the graph represent the energy consumption data features at a certain time, and the edges represent the dependency relationship between nodes, i.e. node features.

[0068] Specifically, in step S2, the constructed spatio-temporal object multi-element fusion model includes a GNN-Transformer model, the GNN-Transformer model includes a GNN model, a Transformer model and a classifier, the GNN model includes a time GNN model with feature concatenation, a spatial GNN model, an environmental GNN model and a device GNN model, wherein:

[0069] The time GNN model includes an input layer, an intermediate layer and an output layer, the time and energy consumption value dimensions are encoded into node features through the input layer, the intermediate layer updates the message and state through residual connection, and the output layer predicts the future energy consumption value, and the update expression at t+1 is:

[0070]

[0071] In the formula: f represents a state update function of a hidden state, which is a global shared function; x e o[v] represents the features of the edges adjacent to node v; is the hidden state of the neighbor node at time t; x n e[v] represents the features of the neighbor nodes of node v;

[0072] The spatial GNN model is similar to the time GNN model, except that the spatial size and energy consumption value dimensions are encoded into node features through the input layer, wherein the energy consumption value dimension is set according to the spatial feature;

[0073] The environmental GNN model is constructed with multiple, similar to the time GNN model, except that the environmental data and energy consumption value dimensions are encoded into node features through the input layer, wherein the energy consumption value dimension is set according to the environmental factor feature;

[0074] The environmental data includes outdoor temperature, outdoor humidity, indoor temperature and indoor humidity, etc.

[0075] The device GNN model is constructed in a manner similar to the time GNN model, except that the attribute data of the device and the energy consumption value dimension are encoded as node features through the input layer, wherein the energy consumption value dimension is set according to the device attribute features;

[0076] The attribute data includes rated power, rated voltage, rated current, and on-off state of the device, etc.

[0077] The time GNN model, the space GNN model, the environment GNN model, and the device GNN model are spliced as the input of the Transformer model.

[0078] The attention weight expression of the Transformer model is as follows:

[0079]

[0080] In the formula, Q represents a query matrix, K represents a key matrix, V represents a value matrix, d k represents the dimension of the key, which is used for scaling to avoid gradient disappearance caused by too large dot product;

[0081] The classifier includes a linear transformation layer and a Softmax layer.

[0082] Specifically, in step S2, the data set is divided into a test set and a validation set, the test set is used for testing, the hyperparameters are adjusted, and the performance and accuracy are evaluated, and the validation set is used for verification.

[0083] Specifically, the specific steps of step S4 include:

[0084] The time GNN model captures the local dynamic dependency relationship between nodes based on time and energy consumption value dimension, and generates node representation containing structural information, i.e., node embedding.

[0085] The space GNN model captures the local dynamic dependency relationship between nodes based on space size and energy consumption value dimension, and generates node representation containing structural information, i.e., node embedding.

[0086] The environment GNN model captures the local dynamic dependency relationship between nodes based on environmental data and energy consumption value dimension, and generates node representation containing structural information, i.e., node embedding.

[0087] The device GNN model captures the local dynamic dependency relationship between nodes based on device attribute data and energy consumption value dimension, and generates node representation containing structural information, i.e., node embedding.

[0088] The node embeddings of the time GNN model, the space GNN model, the environment GNN model, and the device GNN model are spliced and input into the Transformer model.

[0089] The Transformer model captures long-distance dependencies across nodes by a self-attention mechanism, gives dynamic weights to the association of different nodes, and concatenates local feature dependencies into global feature dependencies;

[0090] The linear transformation layer maps the global features output by the Transformer model to the required result dimension;

[0091] The Softmax layer converts the linear output into a probability distribution and outputs the prediction result based on a preset decision rule.

[0092] Specifically, in step S5, the expression of the improved Ludada method is:

[0093]

[0094] In the formula, x represents real-time energy consumption data; represents the mean of real-time energy consumption data within 24 hours; x forecast represents the predicted value, and the weight of the predicted value is increased; a represents a custom parameter; and s represents the standard deviation of real-time energy consumption data within 24 hours;

[0095]

[0096] In the formula, n represents the number of samples; represents the deviation between a certain data and the sample mean.

[0097] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1.A real-time energy consumption anomaly data monitoring method of a spatiotemporal object multi-element fusion model, characterized in that, The method comprises the following steps: S1: collecting energy consumption historical data and preprocessing, saving the preprocessed energy consumption historical data to the database as a data set; S2: constructing a space-time object multi-element fusion model, training through the data set until convergence; S3: collecting real-time energy consumption data; S4: predicting whether there is abnormal data in the real-time energy consumption data based on the converged space-time object multi-element fusion model; S5: inputting the real-time energy consumption data and the space-time object multi-element fusion model prediction result to the improved Ludida method, collecting the energy consumption data within 24 hours, calculating the standard deviation and mean value, and based on the real-time energy consumption data, the space-time object multi-element fusion model prediction result, the energy consumption data mean value within 24 hours and the custom parameters, calculating the residual, if the deviation value is less than three times the standard deviation of the energy consumption data within 24 hours, it is an abnormal value, otherwise, it is a normal value, and outputting the identification result of the real-time energy consumption data. 2.The method of claim 1, wherein the method further comprises: determining the abnormal energy consumption data of the space-time object based on the multi-element fusion model. In the step S1, the energy consumption historical data includes time, space size, environment, device attribute and energy consumption value dimensions. 3.The method of claim 2, wherein the method further comprises: determining the abnormal energy consumption data of the space-time object based on the multi-element fusion model. In the step S1, the energy consumption historical data preprocessing comprises the following steps: Cleaning the collected energy consumption historical data to remove abnormal data noise, and supplementing missing data through spline interpolation processing; Unifying the energy consumption historical data according to international standard units; Normalizing the energy consumption historical data according to the maximum energy consumption calculation result; Extracting the features of the energy consumption historical data and representing them as fixed-size vectors; Constructing a directed acyclic graph according to the time sequence of the energy consumption historical data features represented by fixed-size vectors, the nodes in the graph represent the energy consumption data features at a certain time, and the edges represent the dependency relationship between the nodes, i.e. the node features. 4.The method of claim 3, wherein the method further comprises: In the step S2, the constructed space-time object multi-element fusion model comprises a GNN-Transformer model, the GNN-Transformer model comprises a GNN model, a Transformer model and a classifier, the GNN model comprises a time GNN model, a space GNN model, an environment GNN model and a device GNN model, wherein: Encoding the time and energy consumption value dimensions as the node features of the time GNN model, and the t+1 time update expression of the time GNN model is: wherein: f denotes a state update function of the hidden state as a globally shared function; x e o[v] denotes a feature of an edge adjacent to node v; is a hidden state of a neighbor node at time t; x n e[v] denotes a feature of a neighbor node of node v; Encoding the space size and energy consumption value dimensions as the node features of the space GNN model by setting the energy consumption value dimension through the space feature, and the others are the same as the time GNN model; Encoding the environment data and energy consumption value dimensions as the node features of the environment GNN model by setting the energy consumption value dimension through the environment factor feature, and the others are the same as the time GNN model; Encoding the device attribute data and energy consumption value dimensions as the node features of the device GNN model by setting the energy consumption value dimension through the device attribute feature, and the others are the same as the time GNN model; The attention weight expression of the Transformer model is: In the formula, Q represents a query matrix; K represents a key matrix; V represents a value matrix; d k represents the dimension of the key, used for scaling to avoid gradient disappearance caused by too large dot product. 5.The method of claim 4, wherein the method further comprises: determining the abnormal energy consumption data of the space-time object based on the multi-element fusion model. In the step S2, the data set is divided into a test set and a validation set, the test set is used for testing, the hyperparameters are adjusted, and the performance and accuracy are evaluated, and the validation set is used for verification. 6.The method of claim 5, wherein the method further comprises: determining the abnormal energy consumption data of the space-time object based on the multi-element fusion model. The specific steps of the step S4 comprise: The time GNN model is based on time and energy consumption value dimensions, captures local dynamic dependency between nodes, and generates node representation, i.e., node embedding, containing structural information; The space GNN model is based on space size and energy consumption value dimensions, captures local dynamic dependency between nodes, and generates node representation, i.e., node embedding, containing structural information; The environment GNN model is based on environment data and energy consumption value dimensions, captures local dynamic dependency between nodes, and generates node representation, i.e., node embedding, containing structural information; The device GNN model is based on device attribute data and energy consumption value dimensions, captures local dynamic dependency between nodes, and generates node representation, i.e., node embedding, containing structural information; The node embeddings of the time GNN model, the space GNN model, the environment GNN model and the device GNN model are spliced and input into a Transformer model; The Transformer model captures long-distance dependency across nodes through a self-attention mechanism, gives dynamic weights to the association of different nodes, and concatenates local feature dependency into global feature dependency; A linear transformation layer maps the global features output by the Transformer model to the required result dimension; A Softmax layer converts the linear output into a probability distribution and outputs a prediction result based on a preset decision rule. 7.The method of claim 6, wherein the method further comprises: determining the abnormal energy consumption data of the space-time object based on the multi-element fusion model. In the step S5, the expression of the improved Rayleigh method is: wherein: x represents real-time energy consumption data; represents the mean of real-time energy consumption data over 24 hours; x forecast represents the predicted value, the weight of the predicted value is increased; a represents a self-defined parameter; s represents the standard deviation of real-time energy consumption data over 24 hours; where: n represents the number of samples; represents the deviation of a certain data from the sample average.

Images