Power grid data stream connection method based on deep learning and parallel computing

Abstract

The present invention discloses a method for connecting power grid data streams based on deep learning and parallel computing, and proposes an efficient data parallel similarity connection strategy. Based on a deep recurrent neural network, they are embedded in a space of a certain dimension, avoiding computing bottlenecks caused by problems such as inconsistent sampling rate or length; at the same time, the representation results learned through the network can successfully capture the spatiotemporal characteristics of the data and improve the model's resistance to noise points; the present invention uses a matrix model to perform distributed calculations on the similarity of vector groups. When represented by vectors, the similarity calculation method is generally the cosine similarity of sequence data. Therefore, the present invention does not need to consider the problem of point alignment and can maintain system stability and connection calculation accuracy when performing sequential connection operations.

Description

Technical Field

[0001] The present invention relates to power grid data, and in particular to a power grid data stream connection method based on deep learning and parallel computing. Background Art

[0002] Power grid data is a typical example of time series data. Similarity-based joins within power grid systems provide fundamental support for various applications, as this operation not only enables transportation and distribution but also plays a crucial role in voltage regulation and grid management. In recent years, with the advancement of related technologies and the widespread application of deep learning, many sequential data types can be transformed into time series with d-dimensional features through embedding representations and deep learning. The general process framework for time series data is fully applicable to power grid data. Time series data is organized in chronological order, with "similarity" serving as the predicate for various basic operations. Numerous distance functions exist for measuring time series similarity, such as Hausdorff distance, DTW distance, LCSS distance, ERP distance, and EDR distance. These methods can capture spatial, structural, and dynamic characteristics of the data series. However, these pairwise distance functions require alignment of partial or complete points, resulting in considerable computational complexity.

[0003] The point matching method for calculating the similarity between two time series will encounter performance bottlenecks in the following three situations:

[0004] 1. When the series state is relatively variable. Time series are rich and complex structures with significant uneven distribution in both time and space. For example, the density and complexity of spatiotemporal data in the central region are quite high, while the number of points in the peripheral regions is significantly sparse. This indicates a serious data skew problem in the spatial distribution.

[0005] 2. The second situation is poor time series data quality, resulting in unreliable calculation results. Because sampling of target objects isn't completely real-time, sequence disorder often occurs. For example, data from remote areas or areas with weak signals can be uneven when calculating distances. Furthermore, sampling records can be affected by other activities. Furthermore, when data arrives at the system, the time is unknown and disordered—this is the classic temporal disorder problem. Due to this phenomenon, similarity can be completely different, as many metrics are based on point-to-point matching.

[0006] 3. The third situation is when the sample points are noisy. If the time series contains many noisy points, it is difficult to overcome the positional errors of the matching points. Traditional similarity calculation functions require alignment of the sample points in the sequence. Noise can significantly impact the distance calculation results. Power grid data is subject to various signal interferences and voltage instability, resulting in a high level of noise in the data series. Failure to properly clean and identify this noise during processing can lead to significant deviations in the results. Summary of the Invention

[0007] Purpose of the invention: The purpose of the present invention is to provide a power grid data stream connection method based on deep learning and parallel computing, so as to maintain the stability of the system and the accuracy of the connection calculation when performing sequential connection operations.

[0008] Technical Solution: This invention describes a method for connecting power grid data streams based on deep learning and parallel computing. The principle is to embed data sequences into a fixed dimensional space, compressing and normalizing the raw data of varying lengths and sampling rates. All embedded vectors are then randomly partitioned by a manager to perform parallel computation of the subsequent matrix model. Each process node calculates vector similarity results on the received time representation and reaggregates them in downstream applications.

[0009] A method for connecting power grid data streams based on deep learning and parallel computing includes the following steps:

[0010] (1) Receive the power grid data stream. For each sequence, it is iteratively input into the recurrent neural network according to the time step and processed by two activation functions;

[0011] (2) processing the sequence received in step (1) in real time, performing embedding representation and compressing the data stream within a time window into a multi-dimensional vector within a certain dimension;

[0012] (3) All processing nodes are logically organized in a matrix pattern to receive the vector sequence from step (2);

[0013] (4) After the vector sequences are fed into the processing nodes in different directions based on step (3), the vector similarity is calculated and normalized;

[0014] (5) Based on the similarity results between different vector sequences obtained in step (4), the similarity connection semantics of the power grid data stream can be expressed;

[0015] (6) The flow connection results obtained in step (5) will be used as a quantitative representation of the correlation of the power grid data flow, and the similarities between the two will be aggregated in step (8);

[0016] (7) The similarity connection of all single processing nodes is completed;

[0017] (8) Aggregation of connection operations performed downstream by the control node;

[0018] (9) Analyze the correlation of power grid data streams based on the aggregation results and obtain the final analysis results for subsequent applications;

[0019] (10) Iteratively process the continuously flowing power grid data.

[0020] This method uses a recurrent neural network with a gating mechanism for data modeling and organizes all processing nodes through a matrix model. Specifically, this gating mechanism is used to address the exploding or vanishing gradient issues that can occur during neural network training. It also effectively reduces computational load and provides robustness against noise and abnormal fluctuations in power grid data. Furthermore, the matrix model's scale is dynamically updated and adjusted based on traffic flow, including row and column expansion and merging operations.

[0021] A computer storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned power grid data stream connection method based on deep learning and parallel computing.

[0022] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for connecting power grid data streams based on deep learning and parallel computing is implemented.

[0023] Beneficial Effects: Compared with existing technologies, this invention offers the following advantages: It uses deep learning methods to extract and represent features from power grid data, effectively compressing massive amounts of data. It can embed sequence data with inconsistent quality, difficult sampling rates, and length alignment into a fixed space, ensuring uniform data representation and providing a comprehensive prerequisite for subsequent similarity calculations. Furthermore, the logical organization of working nodes into a matrix effectively reduces workload and allows for dynamic adjustments based on grid traffic. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flow chart of the steps of the present invention;

[0025] Figure 2 Build a schematic diagram of deep representation learning and parallel computing architecture;

[0026] Figure 3 To adopt a recursive neural network graph based on gate units;

[0027] Figure 4 To adopt a distributed computing graph based on a matrix model;

[0028] Figure 5 is an example of a power grid data sequence. DETAILED DESCRIPTION

[0029] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0030] like Figures 1-2 As shown, a method for connecting power grid data streams based on deep learning and parallel computing includes the following steps:

[0031] (1) Receive power grid data stream;

[0032] (2) Figure 3 As shown, the sequence received in step (1) is processed in real time, embedded representation is performed and the data stream within a time window is compressed into a vector within a certain dimension;

[0033] (3) Figure 4 As shown, all processing nodes are logically organized in a matrix mode, and receive the vector sequence from step (2) respectively;

[0034] (4) After the vector sequences are fed into the processing nodes in different directions based on step (3), the vector similarity is calculated and normalized;

[0035] (5) Based on the similarity results between different vector sequences obtained in step (4), the similarity connection semantics of the power grid data stream can be expressed;

[0036] (6) The flow connection results obtained in step (5) will be used as a quantitative representation of the correlation of the power grid data flow, and the similarities between the two will be aggregated in step (8);

[0037] (7) The similarity connection of all single processing nodes is completed;

[0038] (8) Aggregation of connection operations performed downstream by the control node;

[0039] (9) Analyze the correlation of power grid data streams based on the aggregation results and obtain the final analysis results for subsequent applications;

[0040] (10) Iterative processing of the continuously flowing power grid data, such as Figure 5 shown.

[0041] Figure 1The flowchart of the steps of the present invention shows that after the power grid data flows into the system, it is firstly subjected to feature extraction through the embedded characterization module (see steps (1) to (2) for details), then through the distributed computing module, in which all processing nodes are organized into a matrix at the logical layer, which can efficiently process data stream connection operations (see steps (3) to (6) for details), and finally through the result output module for data stream aggregation and output (see steps (7) to (9) for details).

[0042] Figure 3 In order to adopt a recursive neural network graph with gate units, the present invention adopts the GRU model for representation learning, and outputs h at each time step. t The hidden layer represents the result.

[0043] Figure 4 In order to adopt a distributed computing graph based on a matrix model, the present invention performs a natural Cartesian product operation on the sequence data flowing in horizontally and vertically through the matrix, effectively reducing the computing load.

[0044] Figure 5 is an example of a power grid data sequence, Represents the jth time series (denoted as the i-th point at the i-th time step. Taking the Euclidean distance as an example, the correlation between T1 and T2 is represented by the distance between the time series.

Claims

1. A method for connecting power grid data streams based on deep learning and parallel computing, characterized in that: The following steps are involved: (1) Receive power grid data stream; (2) Real-time processing of the data stream received in step (1), embedding representation and compressing the data stream within a time window into a vector within a certain dimension; (3) All processing nodes are logically organized in a matrix pattern to receive the vector sequence from step (2) respectively; (4) After the vector sequences are fed into the processing nodes in different directions based on step (3), the vector similarity is calculated and normalized; (5) Based on the similarity results between different vector sequences obtained in step (4), the similarity connection semantics of the power grid data stream can be expressed; (6) The similarity results obtained in step (5) will be used as a quantitative representation of the correlation of the power grid data flow, and the similarities between the two will be aggregated in step (8); (7) The similarity connection of all single processing nodes is completed; (8) Aggregation of connection operations performed downstream by the control node; (9) Analyze the correlation of power grid data streams based on the aggregation results and obtain the final analysis results for subsequent applications; (10) Iteratively process the continuously flowing power grid data; Among them, a recurrent neural network with a gating mechanism is used for data modeling, and all processing nodes are organized through a matrix model; at the same time, the specific scale of the matrix model will be dynamically updated and adjusted according to the specific traffic, including row and column expansion or merging operations.

2. A computer storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the power grid data stream connection method based on deep learning and parallel computing as described in claim 1.

3. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the power grid data stream connection method based on deep learning and parallel computing as described in claim 1.

Images