Digital Twin Modeling for Hydropower System Based on Radio Frequency Identification Data Collection

Abstract

The safe and steady operation of hydropower generation systems is crucial for electricity output in the grid. However, hydropower stations have complicated interior structures, making defect detection difficult without disassembly inspections. The application of digital modeling to hydropower stations will effectively promote the intelligent transformation of hydropower stations as well as reduce the maintenance costs of the system. This study provides a model of the power generating and transmission system for hydropower plants, with an emphasis on primary equipment and measured data. The model utilizes PSCAD to digitalize state response in hydropower plants with various short-circuit faults. The fault information is identified and learned using the Adaptive Time–Frequency Memory (AD-TFM) deep learning model. It is demonstrated that our proposed method can effectively obtain the fault information through radio frequency identification (RFID). The accuracy of the traditional method is 0.90, while the results for AD-TFM show a fault classification accuracy of 0.92, which is more than enough to identify multiple fault types compared to the existing methods.

1. Introduction

As the economy and renewable energy grow, the power supply system has to be more stable [1]. Hydropower plays an important role in China’s electric power industry; its stable operation is vital to both the national economy and security [2]. Hydropower units, a key component of hydropower facilities, are becoming increasingly complex and sophisticated as hydropower production scales up [3,4]. The increasing complexity of hydropower systems has brought great challenges to the traditional maintenance and troubleshooting processes. traditional maintenance modes are based on post-maintenance and repetitive routines [5]. Expanding the scale of hydropower production and improving the quality of hydropower operation are the structural bases needed to support the long-term development of renewable energy strategies. Among them, ensuring the continuous and stable operation of hydropower systems is the most basic and important task [6]. However, post-maintenance can only address equipment that has already had incidents. The economic losses that have already occurred are unavoidable. On the other hand, scheduled routine maintenance might lead to over-maintenance owing to a lack of knowledge about the equipment’s current status. Both maintenance strategies do not support the stable functioning of hydropower systems. Therefore, the promotion of intelligent operation and maintenance of hydropower stations, the realization of online monitoring, and the prevention of system failure in hydropower stations to avoid large-scale accidents has become an important issue in the development of the hydropower industry in China.

Digital twins first evolved from the concept of a “virtual digital expression equivalent to physical product” proposed by Grieves [7], and have attracted wide attention in recent years. A digital twin is the mapping of physical-space entities to the digital space and the simulation and characterization of the attributes, behaviors, rules, etc. of real physical entities in the environment through a multi-dimensional, multi-scale, multi-disciplinary, and multi-physical-quantity dynamic virtual model of physical entities in a cyber space [8,9]. In an ideal state, the digital twin and its corresponding physical model show the same performance in terms of geometric characteristics, physical characteristics, and application characteristics, and the real operation and state of the physical entity can be reflected through the simulation of the digital twin model [10,11]. Digital twins can control and execute tasks with physical entities through real-time simulation and optimization of virtual digital bodies, which can be transmitted to technicians for fault analysis and decision-making [12,13].

Through the construction of a digital model of a hydropower station system, the real data in the hydropower station system is mapped to a cyber hydropower station to simulate its operating state. As a result, intelligent perception, real-time interaction, and artificial intelligence algorithms can be realized. This provides for greater development of the intelligent operation and intelligent maintenance of the primary equipment and system in the hydropower station. The digital twin serves as a pivotal tool for operational optimization and longevity. For instance, the world’s first structural digital twin of a hydroelectric power station was created for the Turlough Hill hydroelectric power station in Ireland. The digital twin was based on advanced algorithms developed over 15 years at the Massachusetts Institute of Technology (MIT), which allowed for real-time, continuous monitoring of large assets [14]. Digital twins of hydropower facilities can optimize their operation, maintenance, and overall performance. Performance monitoring and analysis through real-time data collection enables operators to identify deviations from optimal conditions and implement corrective measures promptly. The application of digital twins in this sector is a testament to the concept’s evolution from a theoretical model to a practical tool that drives efficiency and sustainability in energy production [15]. The summary of the literature review of Digital Twin model is given in

Table 1. Summary of the literature review.

Methodology	Key Finding	References
Adaptive learning with recursive least squares algorithm	Developed dynamic models for digital twins that accurately mimic hydropower turbine dynamics	[16]
Systematic literature review and bibliometric–qualitative analysis	Highlights the need for integrated information systems in digital twins for watershed management	[15]
Rigorous method	Provides a comprehensive overview of digital twin applications, aiding in the identification of best practices	[14]
Virtual replica	Highlights the use of digital twins for design, production, prognostics, and health management	[17]

.

Fault diagnosis in power systems is a critical process for maintaining system reliability, safety, and efficiency. In recent years, the design of radio frequency identification (RFID) sensor tags has garnered significant interest. The backscattering mechanism underlying their operation endows RFID sensor tags with a simple circuit structure, low cost, and low power consumption, making them particularly well-suited for long-term monitoring [18]. Each RFID sensor tag possesses unique ID information, facilitating rapid and precise fault location. Previous work [19] has explored the design and implementation of an RFID-based fault diagnosis system, the integration of RFID sensor tags with existing power system infrastructure, and the methodologies for real-time data acquisition, processing, and fault detection.

By leveraging RFID technology, this paper aims to provide a robust, cost-effective, and energy-efficient solution for continuous monitoring and timely fault diagnosis, thereby enhancing the reliability, safety, and efficiency of the power system.

• Due to the increasing complexity of power systems and the huge amount of monitoring data with very few malfunctions, digital twins of hydropower equipment are a challenging topic. With the development of artificial intelligence, data-driven intelligent algorithms have been widely used in the fault detection of power systems [20]. Since power system fault information is positively correlated with time series, long short-term memory recurrent neural networks (LSTMs), which use contextual time series and shows good performance in time series prediction and classification, are applied to power system fault identification. This technique has also achieved good performance in time series estimation and classification [21,22]. Due to the lack of frequency domain analysis of LSTM, fault signals are non-stationary, and the fault data samples are scarce. A fault identification scheme based on LSTM cannot completely extract the features of non-stationary initial faults [23]. In order to solve the above problems as existing in LSTM, researchers have embedded adaptive wavelet transform into LSTM neurons to realize the transformation of time-domain signals to frequency-domain signals, and then proposed adaptive Time–Frequency Memory (AD-TFM) units. Based on the AD-TFM unit, the AD-TFM-AT model of a neural network with an attention mechanism is designed, and the AD-TFM-AT neural network can effectively classify the fault messages of a time series in power system faults after the training of data sets [24]. Our contributions are summarized as follows: digital twin technology is used to model and reconstruct the hydropower station’s primary equipment as well as the whole system.
• An optimization framework for the intelligent management and maintenance of hydropower stations as well as online monitoring is provided by the AD-TFM-AT neural network, which is utilized to perform unsupervised learning and fault identification of multi-dimensional fault signals.
• RFID tags are used to collect the data in order to simulate the operations of a hydropower station’s digital system to collect multi-dimensional fault signal information.

2. Digital Twin Model Architecture and Building

2.1. Digital Twin Model Architecture

The architecture of the digital twin model consists of four layers according to their functions and characteristics, and the architecture of the digital twin model is shown in Figure 1:

• Client Use Domain: This layer facilitates the client’s interaction with the system, providing a user-friendly interface for monitoring and control.
• Digital Twin Layer: The core component responsible for modeling and data processing. It is divided into two sub-layers:

  ○

  Modeling: Handles virtualization and model management of the hydropower station.

  ○

  Service: Utilizes model analysis to address practical problems and improve decision-making.

• Interactive Layer: Manages the upload of measurement data from the hydropower station and the transmission of control instructions.
• Hydropower Station Layer: Represents the operational state of the hydropower station and interacts with the upper layers to ensure seamless data flow and operational control.

This paper is based on a digital twin model and the data relationship between the hydropower station and the virtual model. The virtual model is run to obtain the twin data, the transfer learning method is used to transfer the twin data to the hydropower station for AD-TFM-AT, and hydropower station fault prediction is realized.

The digital twin model extracts the core data of the digital twin for separate analysis, extracts the hydropower station data from the operation process of the hydropower station for analysis, establishes a twin model to extract the twin data for analysis, uses the characteristics of the hydropower station data to test the modeling effect, and uses the twin data to conduct algorithm experiments on the hydropower station data. The goal is to use twin data to predict the state of the hydropower station.

2.2. Digital Twin Model Building

The digital twin model of power generation and transmission in hydropower stations is built with PSCAD software. The digital twin model undergoes rigorous simulation testing within PSCAD, where its outputs are compared with real-world data from the hydropower station, and key performance indicators such as accuracy, precision, and recall are measured. Cross-validation techniques are used, where the dataset is divided into training and testing subsets to comprehensively evaluate the model’s performance. RFID technology is used for real-time monitoring, allowing for continuous integration of new data and ongoing adjustments to the digital twin model. This real-time approach ensures that the model remains accurate over time by refining its predictions based on up-to-date information. The digital twin model has been practically applied at the Yangxia 2 hydropower plant, demonstrating high accuracy in fault prediction (92%) and a 10–15% increase in operational efficiency through optimized parameter settings. These verification methods confirm the model’s robustness and practical utility in managing and predicting the operational state of hydropower stations. The single-line diagram of the digital twin model built in PSCAD is shown in Figure 2.

The model mainly includes the Yangxia 2 power plant, generator module, three-phase power transformers module, transmission lines module, and flexible HVDC modules composed of a modular multilevel converter (MMC) and circuit protection breaker device. The system frequency of the digital twin model of the whole hydropower station is set at 50 Hz, the output power of the generator is 5.2 MVA, the rated power of the three-phase transformer is 475 MVA, the length of the transmission line is 30 km, and the voltage when the transmission voltage reaches the MMC is 230 KV. The specific model of the generator module is shown in Figure 3a. The generator model consists of a permanent magnet generator, two-level 5 MW VSC, and external AC source. Figure 3b shows the modular multilevel converter (MMC VSC) in the system model. The MMC VSC consists of external AC source, two VSC models, and two MMC PMW models. The rated power of the generator is 5.2 MVA, and the power of the two-level Voltage Source Converter (VCS) connected to the generator is 5.0 MW. On this basis, the radio frequency identification (RFID) technology is used to collect the actual data of the Jinchuan Hydropower Station unit of the State Energy Group Dadu River Basin Hydropower Development Co., Ltd, Sichuan, China, using the RFID tag data as shown in Figure 3 [25], and then the various parameters of the PSCAD model are refitted and the digital twin is realized through the updated PSCAD model.

After the system model is built, in order to simulate a short-circuit fault during the operation of the hydropower station, three kinds of faults, namely, three-phase short-circuit fault, single-phase short-circuit fault, and four-phase short-circuit fault, are set at the fault points (F1), F2, F3, F4, and F5 as shown in Figure 2 to simulate the system response state of the hydropower station system when a sudden short-circuit fault occurs during actual operation. In the process of fault simulation, the initial time of the short circuit fault is set to 3 s, the duration of short circuit fault is set to 0.15 s, and the simulation running time of the whole hydropower station model is 5 s. After the fault simulation, the data of the running system was obtained from each data measurement point through the RFID in the model as shown in Figure 2.

According to Figure 4, the REs values for faults in the hydropower station’s digital twin model of RFID 3, 4, 5, 6, 7, and 8 all fall within the range of [0, 1.7]. In contrast, the RE values for RFID 1 and 2 significantly exceed this range under no fault, single-phase faults, and three-phase faults. As stated in [26], the RE value is influenced by the fault’s location; therefore, it can be inferred that the winding fault is located near RFID 1 and 2. Additionally, the higher RE value of RFID 1 compared to RFID 2 indicates that the fault is closer to RFID 1.

Figure 5 shows the active power data curve of trouble-free operation and the active power curve sampled by the measuring point at the output end of the generator when a single-phase fault and three-phase fault occur. Figure 6 shows the reactive power data curve of trouble-free operation and the reactive power curve sampled by the measuring point at the output end of the generator when a single-phase fault and three-phase fault occur. It can be obviously seen that, compared with no fault, the active power of a single-phase short-circuit fault decreases to below 60 MVA, and the reactive power rises rapidly to 40 kW, which changes significantly compared with normal operation; the active power of the system decreases to a negative value instantaneously when the three-phase short-circuit fault occurs, while the reactive power increases rapidly. After the fault lasting 0.15 s ended, the system recovered from the transient electromagnetic imbalance to a stable state.

3. Intelligent Fault Identification

The intelligent fault identification consist of three parts: data collection, construction of AD_TFM, and preprocessing of fault data. The whole procedure of fault diagnosis is shown in Figure 7.

3.1. Information Collection Principles of RFID

RFID technology operates through an intricate system comprising electronic tags, readers, and antennae to facilitate real-time data collection and analysis [27]. When a fault occurs in a hydropower station, RFID readers, equipped with antennae, emit a signal that activates the RFID tags positioned at the fault points. The activated tags then transmit their stored data, which includes critical parameters such as voltage and current readings, back to the readers. The RFID readers, through their antennae, capture this data and relay it to a centralized data processing system. The processing system aggregates and analyzes the collected data, providing a comprehensive overview of the hydropower station’s response to each type of fault. The real-time data collection enabled by RFID technology allows for precise monitoring of the fault dynamics, facilitating validation and refinement. The established deep learning model is then used to reconstruct the vibration data of each RFID tag, such that the RE is then obtained.

3.2. Construction of AD-TFM Neurons

In this paper, the AD-TFM neuron structure formed by embedding wavelet transform into LSTM is shown in Figure 8. It is composed of a joint forgetting gate, an input gate, and an output gate. The joint forgetting gate determines the hidden state of the previous time series and the current time series, and the input gate selects the data to be updated. In the unit information update part, the sequence signal after adaptive wavelet transformation and the information retained by the joint forgetting gate are updated, and the updated cell state is processed by the output gate to obtain the hidden state in the current time series [28]. The construction of AD_TFM neurons consists of four layers:

3.2.1. Joint Forgetting Gate

The joint forgetting gate consists of three parts, namely, frequency forgetting gate $f_t^{fre}$, time forgetting gate $f_t^{tim}$, and state forgetting gate $f_t^{ste}$. The joint forgetting gate divides input information and hidden state information into a K-dimensional time domain, J-dimensional frequency domain, and D-dimensional state domain, respectively. Where $W_{*}$ and $U_{*}$ are the weight matrix, $b_{*}$ is the bias matrix, and $h_{t-1}$ is the hidden state:

$$f_t^{fre}=sigmoid\left(W_{ste}{x}_t+U_{ste}{h}_{t-1}+b_{ste}\right),$$

(1)

$$f_t^{tim}=sigmoid\left(W_{tim}{x}_t+U_{tim}{h}_{t-1}+b_{tim}\right),$$

(2)

$$f_t^{fre}=sigmoid\left(W_{fre}{x}_t+U_{fre}{h}_{t-1}+b_{fre}\right),$$

(3)

3.2.2. Neuron Input Gate

The AD-TFM neuron input gate is shown in the following formula:

$$i_t=sigmoid\left(W_i{x}_t+U_i{h}_{t-1}+b_i\right),$$

(4)

$$g_t=tanh(W_g{x}_t+U_g{h}_{t-1}+b_g),$$

(5)

$$i{g}_t=i_t·g_t,$$

(6)

3.2.3. Wavelet Transformation

By adding wavelet transform to the input gate in AD-TFM, the time-domain signal is converted to the frequency-domain signal, and the wavelet change formula is as follows:

$${\psi }_{k,j}=exp(i\cdot \frac{{\omega }_0}{a}\cdot \left(\frac{t+b}{2^j}-k\right))\cdot exp(\left(-\frac{1}{a}\right)\cdot \left(\frac{t+b}{2}-k{)}^2\right),$$

(7)

$$a=tanh(W_{a}i{g}_t+b_a),$$

(8)

$$b=tanh(W_{b}i{g}_t+b_b),$$

(9)

3.2.4. Output Layer

The output layer formula is as follows:

$$O_t=sigmoid\left(W_o{x}_t+U_o{h}_{t-1}+b_o\right),$$

(10)

3.3. Preprocessing of Fault Data

In the development of the AD-TFM neural network, the selection of the time window length and the interval (T) is critical to the model’s accuracy. The time window length of 100 was chosen based on preliminary tests that indicated it provided a balance between computational efficiency and sufficient data granularity to capture the dynamics of faults within the hydropower station system. This length ensures that each window contains enough data points to represent the fault characteristics without being so large as to dilute the fault signature with non-fault-related data.

In order to ensure the accuracy of the AD-TFM neural network, it is necessary to extract and encode a large number of data sets to form the input of the model. After simulating three types of faults in the hydropower station system, the active power and reactive power data at different measurement points of the line and the three-phase voltage at the bus bar are obtained as the fault data sources. The time window sliding method [24] as shown in Figure 9 is adopted for the original fault data. After obtaining the data of a window, the sampling window slides backward for a certain interval T for cyclic data collection until the end, and window sampling is carried out to increase the amount of fault data. The start time of the whole data sampling is 2.9 s, and the end time of sampling is 4 s. A total of 960 sets of data are obtained as the original data of the fault.

The obtained fault data is normalized as shown in Formula (11) to meet the requirements of reasonable data distribution [26]. In the formula, the normalized data is the original fault data. The normalized data set was reshaped into a (96,010,012) sample data set, and the data was integrated into training and test sets according to a 7:3 ratio division.

$$y\prime =\frac{y-min(y)}{max(y)-min(y)},$$

(11)

The label of the data set is processed by one-hot coding [29]. The coding method is shown in

Table 2. Fault label coding.

Fault Category	Fault One-Hot Label
Normal	1000
Single-phase short-circuit fault	0100
Two-phase short-circuit fault	0010
Three-phase short-circuit fault	0001

. The activation value of one-hot coding is 1, indicating that the input data corresponds to the fault category, and the activation value is 0, indicating that the input data does not correspond to the fault type.

4. Training Results of AD-TFM Neural Network

The established AD-TFM neural network is set as output category 4, training time 150, training batch size 16, learning rate 0.001, feature parameter dimension 12, and the other parameters as shown in the literature. After training of the AD-TFM neural network, AD-TFM is tested with the test set to evaluate its performance for fault classification. In order to reflect the performance of the AD-TFM neural network, accuracy, recall rate, F1-score, and ROC curve are used as indicators to measure the fault category of the AD-TFM. The accuracy is the ratio of the number of correct predictions to the total number. After testing, AD-TFM has an accuracy of 0.92 for fault classification of 289 test data points.

The accuracy, recall rate, and F1-score of the AD-TFM neural network for four kinds of fault classification are shown in

Table 3. Fault classification performance.

Fault Category	Accuracy	Recall Rate	F1-Score	Sample Size
Normal	0.88	1.00	0.93	28
Single-phase short-circuit fault	1.00	0.91	0.95	79
Two-phase short-circuit fault	0.97	0.81	0.89	86
Three-phase short-circuit fault	0.84	0.99	0.91	95

. The accuracy is the ratio between the number of samples correctly classified as faults and the number of samples classified as faults in the whole population. The accuracy of AD-TFM for single-phase short-circuit fault is 1.00 at the highest, and that for three-phase short-circuit fault identification is 0.84 at the lowest. The recall rate of the AD-TFM network for no fault is 1.00, and the recall rate for dual-phase fault is 0.81. The higher the recall rate, the stronger the ability of the AD-TFM to recognize positive samples. The accuracy and recall rate of F1-score comprehensive courses are calculated. The higher the F1-score value, the better the classification effect of the neural network. As can be seen from Table 2, the highest F1-score of AD-TFM for a single-phase short-circuit fault is 0.95, and the lowest F1-score of AD-TFM for a dual-phase short-circuit fault is 0.89; the study presented in [30] used a cascaded model that leveraged the Random Forest classifier in combination with knowledge reasoning, and the fault accuracy was 90%. It shows that the model has good classification performance for single-phase short-circuit faults, but poor classification performance for dual-phase and three-phase short circuit faults.

In addition to evaluating various parameters, the performance of the fault identification model is assessed using the Receiver Operating Characteristic (ROC) curve, where a larger area under the curve (AUC) signifies better model efficiency. Figure 10 shows the ROC curves for the AD-TMF model in identifying four types of faults. The normal fault category is represented by the green line and denotes the system’s performance under standard, fault-free conditions; SGF is represented by the yellow line, and TGF and MTF, with each fault’s ROC curve having an AUC of 0.99, indicate high classification accuracy and fault identification efficiency. To enhance clarity, Table 2 provides a comparison based on key metrics like accuracy, recall rate, F1-score, and sample size for each fault category. It ensures a robust evaluation of the digital twin model designed. The high AUC values across all fault types confirm that AD-TMF-AT has a high performance for fault identification, effectively distinguishing between different types of faults.

5. Discussion and Limitation

5.1. Discussion

We introduce a digital twin framework tailored to the hydropower system. A key highlight of our study is the significant advancement in RFID technology for fault diagnosis. Our model utilizes an AD_TFM algorithm for fault detection, resulting in a remarkable boost in fault identification accuracy across various fault types. Through extensive testing and validation, our model has consistently yielded accuracy rates surpassing 92%, surpassing contemporary machine learning methodologies.

A significant achievement lies in seamlessly integrating domain expertise into the fault diagnosis methods. Our model’s knowledge repository comprises extensive historical data, industry norms, and rules crafted by experts. This amalgamation guarantees not just fault detection but also furnishes a deeper contextual comprehension of the underlying issues. We contend that this fusion of domain knowledge distinguishes our methodology from purely data-centric approaches, profoundly enhancing the interpretability and dependability of our outcomes.

5.2. Limitation

Constrained by the experimental setup, the fault data discussed in this paper mainly come from simulations conducted on a digital twin model of a hydropower system. Nevertheless, given the robustness and generalization capabilities of the AD_TFM, enhanced performance could be achieved through training with more advanced neural networks. Furthermore, our investigation has thus far concentrated solely on system-wide fault types, neglecting faults specific to generators and the most severe fault scenarios. Addressing these aspects could potentially lead to improved fault detection accuracy.

6. Conclusions

This paper presents a comprehensive model of digital twin modeling for hydropower plants, with a focus on primary equipment and measured data, aiming to enhance detection and intelligent maintenance. Utilizing PSCAD, the study simulates the system response under various short-circuit fault conditions, including single-phase, two-phase, and three-phase faults. Fault information is extracted using a time window sliding approach, creating a dataset for the AD-TFM deep learning model based on long short-term memory recurrent neural networks. The results demonstrate that the AD-TFM model achieves a fault classification accuracy of 92%, surpassing existing methods, particularly in identifying single-phase faults. The AD-TFM algorithm can be used to improve fault identification accuracy and explore its real-time application in online fault detection systems, while also considering the integration of additional fault types and conditions to enhance robustness and applicability in various operational scenarios.