Fault Warning Technology of Comprehensive Energy System Based on Digital Twins

The traditional energy industry has many problems, such as huge losses and non-renewability. With the development of informatization and digitalization, integrated smart energy systems have become a current research hotspot. At present, the research on integrated smart energy is still in the theoretical stage, and the fault warning part for integrated smart energy is still very scarce. Therefore, this article proposes a comprehensive energy failure warning technology based on digital twins. Firstly, the comprehensive energy system based on digital twins is divided into five layers, and the SOM algorithm is used to analyze the system’s status data for fault warning. Finally, the superiority of the proposed method was verified through experiments.


Introduction
Nowadays, the world is facing a sharp increase in energy consumption, which has caused serious environmental pollution problems and is becoming increasingly serious.How to solve the excessive energy consumption in the traditional energy industry and empower the use of emerging green energy has become the hottest topic in the international community [1].Traditional energy systems are usually only single-acting forms of energy, such as electricity, natural gas, and thermal energy.Among them, the utilization rate of energy, the conversion rate of renewable energy, and how to save energy and reduce emissions have always been in a bottleneck period.In recent years, with the advancement of the digital age, the global energy industry has also invested a lot of energy in information construction.The concept of comprehensive intelligent energy systems has gradually penetrated people's hearts.Comprehensive intelligent energy is replacing traditional energy systems and promoting the development of the energy industry [2][3].
Professor Michael Griffiths from the United States proposed the concept of digital twins in a paper [4].The paper indicates that virtual entities and related subsystems can be constructed in the virtual information space to represent physical devices, and bidirectional dynamic relationships can be established with physical device data in real space throughout the entire product lifecycle.From this concept, we can see that the concept of digital twins is reflected in product design, from subsequent production testing to mass production manufacturing and finally to products that serve users.These processes all incorporate the concept of digital twins.NASA was the first country to use digital twin technology in practical applications, and they have achieved significant results in intelligent manufacturing using this technology.Subsequently, countries invested a large amount of manpower and material resources in digital twin technology.Professor Tao Fei's domestic team proposed the concept of a digital twin workshop for the first time, integrating digital twin technology with workshop production to achieve the integration of physics, models, data, and services.
The comprehensive smart energy system mainly focuses on three attributes: physical energy attributes, industrial production attributes, and public service attributes .The future development goal of integrated intelligent energy systems is not only to perfectly solve the energy supply problem of the entire society but also to focus on the development of the entire society.The various information technologies in the era of big data provide convenience for data collection, transmission, and analysis of integrated intelligent energy systems and also provide fertile soil for developing integrated intelligent energy systems [5][6].However, the current data are generated under specific businesses, and there is a lack of connection and relationship between them.There are many obstacles between data, and there are also many information silos.Based on the digital twin, we utilize its ability to combine virtual reality information and its dynamic and real-time features to organically integrate multi-source data generated by intelligent energy systems, analyze and extract features, visualize, and achieve the integrated state presentation and analysis application of energy systems [7].
The state of the integrated intelligent energy system implemented by digital twins has a mixed features of massive, multi-source, complex, dynamic, strong nonlinearity, and uncertainty [8].The integration of multi-source measurements has been completed, system characteristics have been analyzed, and development trends have been predicted.It is necessary to have a deep understanding and professional knowledge of its performance status and characteristics.When using integrated intelligent energy systems, when system equipment malfunctions, it directly affects the operation of the entire power system and the energy supply of the entire process.If system management personnel cannot detect system equipment faults in a timely manner during system operation, it will lead to unpredictable consequences.When faults occur directly in the production environment, they are often difficult to solve, causing huge losses to the system and consuming more manpower, material resources, and financial resources.Therefore, the fault warning of integrated intelligent energy systems will become a research hotspot in the future.The digital twin technology is used to model the system, dynamically present system status, and collect system status.The SOM algorithm is used to achieve fault prediction and provide fault warning for the digital twin model [9].

Overview of the proposed method
The full English name of fault warning is Fault Prognosis or FP for short.It is mainly used to monitor the status data of the equipment and use the data obtained from the monitoring to analyze the probability of failure and give early warning to it, which helps to reduce the cost and prolong the life of the equipment [10].
As shown in Figure 1, the multi-source feature of digital twin technology is used to integrate data into a comprehensive smart energy system.The operation log based on the energy system, equipment data, and statistical information of existing faulty equipment can be used for more comprehensive fault detection.Predict the equipment life and failure probability of the energy supply link and provide a more accurate early warning function.2. A common SOM network has two parts: an input layer and an output layer.The computing layer we often refer to is the place where competition occurs, which is the common output layer.The output layer is a series of neural cells with a one-dimensional or two-dimensional structure, that is to say, the computing layer is topological.It can also be seen that SOM plays a role in dimensionality reduction: high-dimensional input data is output to the output layer through algorithms.In traditional k-means algorithms, we must first specify the value of the number of clusters k.The value of k has a significant impact on the accuracy of the algorithm.In the SOM algorithm, we also need to first determine the geometric relationship between clusters and select topological structures to better cluster neural cells.Neural cells gather around nodes in these topological structures.

Figure 2. Network Topology Diagram
The SOM algorithm converts the high-dimensional input mode into one-dimensional or twodimensional discrete mapping, mainly through topological placement and automatic conversion.Higher-dimensional mapping is possible but not uncommon.In the process of competitive learning, neurons selectively adapt to different input patterns (stimuli) or categories of input patterns.In this way, the nerve cells (that is, the nerve cells of the winner) are arranged in order, creating a coordinate system that is meaningful to the input features on the grid.Based on this, the SOM algorithm will generate the mapping required by the input pattern.
The various stages of the SOM algorithm are shown in Figure 3, which can be summarized as follows: 1) The network structure of SOM is determined.An appropriate number of input layer neural cells and output layer neural cells is selected based on the dimensions of input layer data, and parameters such as neural cell weight vector, network learning rate, and domain function are initialized.
2) The weight vectors of input data and neural cells are normalized.
3) The input data is imported into the SOM network, and the distance between the model and all neural cells in the output layer is calculated to determine the winning neural cell.
4) The winning area is determined according to the radius, and the weight vector of all nerve cells is updated according to the winning area and activation function.
5) The network learning rate and domain function are updated.6) A complete iterative process is completed.If the number of iterations exceeds the set threshold, we skip to Step (7); otherwise, we skip to Step (3).
7) The best performing neural cells for each input data are recorded, and the input data under that neural cell is classified into a category.

The proposed algorithm mapping organization
Firstly, the x point in the input space is mapped to the I (x) point in the output space, as shown in Figure 4:

Self-organization
The self-organization process is as follows: a) Initialization: First, all the connected weight values with a small random value are initialized to obtain an initially connected graph.
b) Competition: Combining with the initially connected graph, for each input mode, the discriminant function value of each nerve cell is calculated through a specific function as a reference for the competition result.We set the nerve cell with the smallest discriminant function value as the winner of the competition.c) Cooperation: The area near the nerve cell that wins the competition is called the topological neighborhood of the excited nerve cell, and the nerve cells in this neighborhood can better cooperate with each other.d) Adaptation: Excited nerve cells can adaptively adjust the relevant connection weights to change the connected graph, thereby reducing the value of the discriminant function related to specific input modes and finally enhancing the winning nerve cells to such input modes.
If the input space is D-dimensional, we can write the input mode as } ,..., , and the connection weight layer between the input unit and the nerve cell in the calculation can be written as } ,..., 1 ; ,..., 1 : , where is the total number of nerve cells.
The square Euclidean distance between the input vector x and the weight vector of each neuron j is defined as the discriminant function, as follows: (1) Through this simple competition between nerve cells, continuous input space can be mapped to the discrete output space of nerve cells through the connected graph.

Experimental verifications
This article refers to many documents and related materials, combined with its own research, and after many experiments, the number of nerve cells in the competition layer is set to 3, the winning field is set to 2, the total number of training iterations of the model is set to 400, and the rates are respectively set to 0.6.Under this parameter setting, the SOM algorithm's fault warning accuracy and accuracy are quite good.The winning value curve of the experiment is shown in Figure 6.It can be seen in Figure 6 that the smart device of the integrated smart energy system maintains a high level of winning value when it is working normally.However, when the smart device has an unexpected situation, its winning value level will be significantly reduced.And by comparing the average value of each feature vector of the smart device fault, it can be shown that the size of the winning value is proportional to the severity of the fault.
At the same time, we use the SOM algorithm to analyze the data envelopment, that is, the measured data of the integrated smart energy equipment is input into the SOM algorithm for training, and the data characteristics obtained by the analysis are used to express the competitive nerve cells.The data envelopment analysis results are shown in Figure 7 and Figure 8. Comparing Figure 7 and Figure 8, it can be seen that the competitive layer neuron 1 in the training data is the best pairing unit under the normal state of the smart device.The accuracy of the good pairing unit has reached 97.60%, which also verifies the effectiveness of the algorithm.

Concluding remarks
This article first introduces the increasingly serious energy problems in the world today and proposes the importance of digitalization to the energy industry.The application of digital and information technology to the integrated smart energy of the energy industry has begun to become a research hotspot.Digital twins and integrated smart energy are introduced.The advantages of using digital twins to solve the integrated smart energy system are listed and proposed using the SOM algorithm to do fault warning for the integrated smart energy system based on digital twins.This article proposes a comprehensive intelligent energy system based on digital twin technology.The entire architecture is divided into five layers, each with its own role.The starting point and technical requirements of fault warning based on this architecture are analyzed.Then the basic principle of the SOM algorithm is introduced, and experimental analysis is conducted.
The widespread application of integrated smart energy systems is of great significance.It not only helps to break through the time and space constraints of the traditional energy industry but also helps to promote the unified planning and overall dispatch of various energy sources, which is a major issue related to people's livelihood.However, the current integrated smart energy system based on digital twins still has problems, such as how to ensure the synchronization of virtual state and system measurement, how to achieve low coupling in the human-computer interaction mechanism, and how to ensure the applicability of digital twin technology.In the future, research on the coordination and optimization of integrated smart energy systems based on digital twins will be a hot direction.

Figure 1 .
Figure 1.Comprehensive Energy Fault Warning Technology Based on Digital Twin System 2.1 SOM algorithm SOM is a clustering algorithm based on neural networks.It is a single-layer neural network.The structure diagram of a common SOM network is shown in Figure2.A common SOM network has two parts: an input layer and an output layer.The computing layer we often refer to is the place where competition occurs, which is the common output layer.The output layer is a series of neural cells with a one-dimensional or two-dimensional structure, that is to say, the computing layer is topological.It can also be seen that SOM plays a role in dimensionality reduction: high-dimensional input data is output to the output layer through algorithms.In traditional k-means algorithms, we must first specify the value of the number of clusters k.The value of k has a significant impact on the accuracy of the algorithm.In the SOM algorithm, we also need to first determine the geometric relationship between clusters and select topological structures to better cluster neural cells.Neural cells gather around nodes in these topological structures.

Figure 4 .
Figure 4. Map DiagramThen each point I in the output space is mapped to the corresponding point w (I) in the input space.As shown in Figure5, each nerve cell is fully connected to all source ganglia in the input layer.A onedimensional map has only one row in the calculation layer.

Figure 7 .
Figure 7. Merit Curve Figure 8. Best Paired Unit Distribution of Test Data